diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..72e8ffc
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1 @@
+*
diff --git a/Dockerfile.training b/Dockerfile.training
new file mode 100644
index 0000000..5a512cb
--- /dev/null
+++ b/Dockerfile.training
@@ -0,0 +1,51 @@
+# cat Dockerfile.training | docker build --build-arg=base=debian:9.8 -t daanzu/kaldi_ag_training .
+# cat Dockerfile.training | docker build --build-arg=base=nvidia/cuda:10.0-cudnn7-devel-ubuntu16.04 --build-arg=cuda=yes -t daanzu/kaldi_ag_training_gpu .
+
+ARG base
+ARG cuda=
+
+FROM $base
+LABEL maintainer="daanzu@gmail.com"
+ARG cuda
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        g++ \
+        make \
+        automake \
+        autoconf \
+        bzip2 \
+        unzip \
+        wget \
+        sox \
+        libtool \
+        git \
+        subversion \
+        python2.7 \
+        python3 \
+        zlib1g-dev \
+        ca-certificates \
+        gfortran \
+        patch \
+        ffmpeg \
+	    vim && \
+    rm -rf /var/lib/apt/lists/*
+
+RUN ln -s /usr/bin/python2.7 /usr/bin/python
+
+RUN git clone --depth 1 https://github.com/daanzu/kaldi-fork-active-grammar /opt/kaldi && \
+    cd /opt/kaldi/tools && \
+    ./extras/install_mkl.sh && \
+    make -j $(nproc) && \
+    cd /opt/kaldi/src && \
+    ./configure --shared ${cuda:+--use-cuda} && \
+    make depend -j $(nproc) && \
+    make -j $(nproc) && \
+    find /opt/kaldi -type f \( -name "*.o" -o -name "*.la" -o -name "*.a" \) -exec rm {} \; && \
+    find /opt/intel -type f -name "*.a" -exec rm {} \; && \
+    find /opt/intel -type f -regex '.*\(_mic\|_thread\|_ilp64\)\.so' -exec rm {} \; && \
+    rm -rf /opt/kaldi/.git
+
+# _mc included for pre-AVX CPUs
+
+WORKDIR /opt/kaldi/
diff --git a/LICENSE b/LICENSE
new file mode 100644
index 0000000..be3f7b2
--- /dev/null
+++ b/LICENSE
@@ -0,0 +1,661 @@
+                    GNU AFFERO GENERAL PUBLIC LICENSE
+                       Version 3, 19 November 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU Affero General Public License is a free, copyleft license for
+software and other kinds of works, specifically designed to ensure
+cooperation with the community in the case of network server software.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+our General Public Licenses are intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  Developers that use our General Public Licenses protect your rights
+with two steps: (1) assert copyright on the software, and (2) offer
+you this License which gives you legal permission to copy, distribute
+and/or modify the software.
+
+  A secondary benefit of defending all users' freedom is that
+improvements made in alternate versions of the program, if they
+receive widespread use, become available for other developers to
+incorporate.  Many developers of free software are heartened and
+encouraged by the resulting cooperation.  However, in the case of
+software used on network servers, this result may fail to come about.
+The GNU General Public License permits making a modified version and
+letting the public access it on a server without ever releasing its
+source code to the public.
+
+  The GNU Affero General Public License is designed specifically to
+ensure that, in such cases, the modified source code becomes available
+to the community.  It requires the operator of a network server to
+provide the source code of the modified version running there to the
+users of that server.  Therefore, public use of a modified version, on
+a publicly accessible server, gives the public access to the source
+code of the modified version.
+
+  An older license, called the Affero General Public License and
+published by Affero, was designed to accomplish similar goals.  This is
+a different license, not a version of the Affero GPL, but Affero has
+released a new version of the Affero GPL which permits relicensing under
+this license.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU Affero General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Remote Network Interaction; Use with the GNU General Public License.
+
+  Notwithstanding any other provision of this License, if you modify the
+Program, your modified version must prominently offer all users
+interacting with it remotely through a computer network (if your version
+supports such interaction) an opportunity to receive the Corresponding
+Source of your version by providing access to the Corresponding Source
+from a network server at no charge, through some standard or customary
+means of facilitating copying of software.  This Corresponding Source
+shall include the Corresponding Source for any work covered by version 3
+of the GNU General Public License that is incorporated pursuant to the
+following paragraph.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the work with which it is combined will remain governed by version
+3 of the GNU General Public License.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU Affero General Public License from time to time.  Such new versions
+will be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU Affero General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU Affero General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU Affero General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If your software can interact with users remotely through a computer
+network, you should also make sure that it provides a way for users to
+get its source.  For example, if your program is a web application, its
+interface could display a "Source" link that leads users to an archive
+of the code.  There are many ways you could offer source, and different
+solutions will be better for different programs; see section 13 for the
+specific requirements.
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU AGPL, see
+<https://www.gnu.org/licenses/>.
diff --git a/README.md b/README.md
new file mode 100644
index 0000000..a2c44a5
--- /dev/null
+++ b/README.md
@@ -0,0 +1,54 @@
+# Kaldi AG Training Setup
+
+[![Donate](https://img.shields.io/badge/donate-GitHub-pink.svg)](https://github.com/sponsors/daanzu)
+[![Donate](https://img.shields.io/badge/donate-Patreon-orange.svg)](https://www.patreon.com/daanzu)
+[![Donate](https://img.shields.io/badge/donate-PayPal-green.svg)](https://paypal.me/daanzu)
+
+Docker image and scripts for training finetuned or completely personal speech models. Particularly for use with [kaldi-active-grammar](https://github.com/daanzu/kaldi-active-grammar).
+
+## Usage
+
+All commands are run in the Docker container as follows. Training on the CPU should work, just much more slowly. To do so, remove the `--runtime=nvidia` and use the image `daanzu/kaldi_ag_training_gpu:2021-08-04` instead the GPU image.
+
+```bash
+docker run -it --rm -v $(pwd):/mnt/input -w /mnt/input --user "$(id -u):$(id -g)" \
+    --runtime=nvidia daanzu/kaldi_ag_training_gpu:2021-08-04 \
+    [command and args...]
+```
+
+Example commands:
+
+```bash
+# Prepare training dataset files
+python3 convert_tsv_to_scp.py -l kaldi_model_daanzu_20200905_1ep-mediumlm-base/dict/lexicon.txt yourdata.tsv [optional output directory]
+
+# Pick only one of the following:
+# Run finetune training, with default settings
+docker run [...] bash run.finetune.sh kaldi_model_daanzu_20200905_1ep-mediumlm-base dataset
+# Run completely personal training, with default settings
+docker run [...] bash run.personal.sh kaldi_model_daanzu_20200905_1ep-mediumlm-base dataset
+
+# When training completes, export trained model
+python3 export_trained_model.py {finetune,personal} [optional output directory]
+```
+
+### Notes
+
+* To run either training, you must have a base model to use as a template. (For finetuning this is also the starting point of the model; for personal it is only a source of basic info.) You can use [this base model](https://github.com/daanzu/kaldi_ag_training/releases/download/v0.1.0/kaldi_model_daanzu_20200905_1ep-mediumlm-base.zip) from this project's release page. Download the zip file and extract it to the root directory of this repo, so the directory `kaldi_model_daanzu_20200905_1ep-mediumlm-base` is here.
+
+* Kaldi requires the training data metadata to be in the SCP format, which is an annoying multi-file format. To convert the standard KaldiAG TSV format to SCP, you can run `python3 convert_tsv_to_scp.py yourdata.tsv dataset` to output SCP format in a new directory `dataset`. You can run these commands within the Docker container, or directly using your own python environment.
+    * Even better, run `python3 convert_tsv_to_scp.py -l kaldi_model_daanzu_20200905_1ep-mediumlm-base/dict/lexicon.txt yourdata.tsv dataset` to filter out utterances containing out-of-vocabulary words. OOV words are not currently well supported by these training scripts.
+
+* The audio data should be 16-bit Signed Integer PCM 1-channel 16kHz WAV files. Note that it needs to be accessible within the Docker container, so it can't be behind a symlink that points outside this repo directory, which is shared with the Docker container.
+
+* There are some directory names you should avoid using in this repo directory, because the scripts will create & use them during training. Avoid: `conf`, `data`, `exp`, `extractor`, `mfcc`, `steps`, `tree_sp`, `utils`.
+
+* Training may use a lot of storage. You may want to locate this directory somewhere with ample room available.
+
+* The training commands (`run.*.sh`) accept many optional parameters. More info later.
+
+    * `--stage n` : Skip to given stage
+
+## License
+
+This project is licensed under the GNU Affero General Public License v3 (AGPL-3.0-or-later). See the [LICENSE file](LICENSE) for details. If this license is problematic for you, please contact me.
diff --git a/cmd.sh b/cmd.sh
new file mode 100644
index 0000000..41cbd6c
--- /dev/null
+++ b/cmd.sh
@@ -0,0 +1,4 @@
+export train_cmd="utils/run.pl"
+export decode_cmd="utils/run.pl"
+export cuda_cmd="utils/run.pl"
+# export cuda_cmd="utils/run.pl -l gpu=1"
diff --git a/convert_tsv_to_scp.py b/convert_tsv_to_scp.py
new file mode 100644
index 0000000..c7adb1e
--- /dev/null
+++ b/convert_tsv_to_scp.py
@@ -0,0 +1,45 @@
+#!/usr/bin/env python
+
+import argparse, os
+
+parser = argparse.ArgumentParser(description='Convert a TSV file to Kaldi SCP files.')
+parser.add_argument('filename', help='The TSV file to convert.')
+parser.add_argument('output_dir', default='dataset', help='The directory to save the output files.')
+parser.add_argument('-l', '--lexicon_file', help='The name of the lexicon file, for filtering out out-of-vocabulary utterances.')
+args = parser.parse_args()
+
+if not os.path.exists(args.filename):
+    raise Exception('File does not exist: %s' % args.filename)
+os.mkdir(args.output_dir, exist_ok=True)
+
+lexicon = set()
+if args.lexicon_file:
+    with open(args.lexicon_file, 'r') as f:
+        for line in f:
+            word, num = line.strip().split()
+            lexicon.add(word)
+
+utt2spk_dict, wav_dict, text_dict = {}, {}, {}
+with open(args.filename, 'r') as f:
+    for line in f:
+        fields = line.strip().split('\t')
+        text = fields[4]
+        wav_path = fields[0]
+        utt_id = os.path.splitext(os.path.basename(wav_path))[0]
+        if lexicon and any([word not in lexicon for word in text.split()]):
+            continue
+        utt2spk_dict[utt_id] = utt_id
+        wav_dict[utt_id] = wav_path
+        text_dict[utt_id] = text
+
+with open(os.path.join(dir, 'utt2spk'), 'w') as f:
+    for (key, val) in utt2spk_dict.items():
+        f.write('%s %s\n' % (key, val))
+with open(os.path.join(dir, 'wav.scp'), 'w') as f:
+    for (key, val) in wav_dict.items():
+        f.write('%s %s\n' % (key, val))
+with open(os.path.join(dir, 'text'), 'w') as f:
+    for (key, val) in text_dict.items():
+        f.write('%s %s\n' % (key, val))
+
+print(f"Wrote training dataset to {args.output_dir}")
diff --git a/export_trained_model.py b/export_trained_model.py
new file mode 100644
index 0000000..186a2d9
--- /dev/null
+++ b/export_trained_model.py
@@ -0,0 +1,34 @@
+#!/usr/bin/env python
+
+import argparse, os, shutil
+
+parser = argparse.ArgumentParser(description='Export trained model.')
+parser.add_argument('type', choices=('personal', 'finetune'), help='Type of trained model.')
+parser.add_argument('output_dir', default='exported_model', help='Directory to save the output model.')
+parser.add_argument('-b', '--base_model_dir', default='kaldi_model_daanzu_20200905_1ep-mediumlm-base', help='Directory of model to copy base files from.')
+args = parser.parse_args()
+
+if not os.path.exists(args.base_model_dir):
+    raise Exception('Base model directory does not exist.')
+if os.path.exists(args.output_dir):
+    raise Exception('Output directory already exists.')
+os.mkdir(args.output_dir, exist_ok=True)
+shutil.copytree(args.base_model_dir, args.output_dir, ignore=shutil.ignore_patterns('dict', 'tree_stuff'))
+os.mkdir(os.path.join(args.output_dir, 'training'), exist_ok=True)
+
+if args.type == 'personal':
+    for name in 'final.mdl tree'.split():
+        shutil.copy2(os.path.join('exp/chain/tdnn1h_sp_online', name), args.output_dir)
+    for name in 'final.dubm final.ie final.mat global_cmvn.stats'.split():
+        shutil.copy2(os.path.join('exp/chain/tdnn1h_sp_online', 'ivector_extractor', name), os.path.join(args.output_dir, 'ivector_extractor'))
+    shutil.copy2('exp/chain/tdnn1h_sp/accuracy.report', os.path.join(args.output_dir, 'training'))
+    shutil.copy2('params.txt', os.path.join(args.output_dir, 'training'))
+
+elif args.type == 'finetune':
+    for name in 'final.mdl'.split():
+        shutil.copy2(os.path.join('exp/nnet3_chain/finetune', name), args.output_dir)
+    shutil.copy2('exp/nnet3_chain/finetune/accuracy.report', os.path.join(args.output_dir, 'training'))
+
+print(f"Wrote exported {args.type} model to {args.output_dir}")
+print("NOTE: You still must run the following in your kaldi-active-grammar python environment:")
+print("python -m kaldi_active_grammar compile_agf_dictation_graph -v -m [model_dir] G.fst")
diff --git a/path.sh b/path.sh
new file mode 100644
index 0000000..c397f99
--- /dev/null
+++ b/path.sh
@@ -0,0 +1,4 @@
+export KALDI_ROOT=/opt/kaldi
+export LD_LIBRARY_PATH="$KALDI_ROOT/tools/openfst-1.3.4/lib:$KALDI_ROOT/src/lib:$LD_LIBRARY_PATH"
+export PATH=$KALDI_ROOT/src/lmbin/:$KALDI_ROOT/../kaldi_lm/:$PWD/utils/:$KALDI_ROOT/src/bin:$KALDI_ROOT/tools/openfst/bin:$KALDI_ROOT/src/fstbin/:$KALDI_ROOT/src/gmmbin/:$KALDI_ROOT/src/featbin/:$KALDI_ROOT/src/lm/:$KALDI_ROOT/src/sgmmbin/:$KALDI_ROOT/src/sgmm2bin/:$KALDI_ROOT/src/fgmmbin/:$KALDI_ROOT/src/latbin/:$KALDI_ROOT/src/nnetbin:$KALDI_ROOT/src/nnet2bin/:$KALDI_ROOT/src/online2bin/:$KALDI_ROOT/src/ivectorbin/:$KALDI_ROOT/src/kwsbin:$KALDI_ROOT/src/nnet3bin:$KALDI_ROOT/src/chainbin:$KALDI_ROOT/src/rnnlmbin:$PWD:$PATH
+export LC_ALL=C
diff --git a/run.finetune.sh b/run.finetune.sh
new file mode 100644
index 0000000..493804d
--- /dev/null
+++ b/run.finetune.sh
@@ -0,0 +1,84 @@
+# docker run -it --rm -v $(pwd):/mnt/input -v $(pwd)/work:/mnt/work -w /mnt/work --user "$(id -u):$(id -g)" daanzu/kaldi_ag_training:2020-11-28 bash run.finetune.sh models/kaldi_model_daanzu_20200905_1ep-mediumlm data/standard2train --num-epochs 5 --train-stage -10 --stage 1
+# docker run -it --rm -v $(pwd):/mnt/input -v $(pwd)/work:/mnt/work -w /mnt/work --user "$(id -u):$(id -g)" --runtime=nvidia daanzu/kaldi_ag_training_gpu:2020-11-28 bash run.finetune.sh models/kaldi_model_daanzu_20200905_1ep-mediumlm data/standard2train --num-epochs 5 --train-stage -10 --stage 1
+
+set -euxo pipefail
+
+nice_cmd="nice ionice -c idle"
+
+[[ $# -ge 2 ]] || exit 1
+
+model=/mnt/input/$1; shift
+dataset=/mnt/input/$1; shift
+
+[[ -d $model ]] || exit 1
+[[ -d $dataset ]] || exit 1
+
+echo "base_model=${model#/mnt/input/}" >> params.txt
+echo "train_dataset=${dataset#/mnt/input/}" >> params.txt
+
+cat <<\EOF > cmd.sh
+export train_cmd="utils/run.pl"
+export decode_cmd="utils/run.pl"
+export cuda_cmd="utils/run.pl"
+# export cuda_cmd="utils/run.pl -l gpu=1"
+EOF
+cat <<\EOF > path.sh
+export KALDI_ROOT=/opt/kaldi
+export LD_LIBRARY_PATH="$KALDI_ROOT/tools/openfst/lib:$KALDI_ROOT/tools/openfst/lib/fst:$KALDI_ROOT/src/lib:$LD_LIBRARY_PATH"
+export PATH=$KALDI_ROOT/src/lmbin/:$KALDI_ROOT/../kaldi_lm/:$PWD/utils/:$KALDI_ROOT/src/bin:$KALDI_ROOT/tools/openfst/bin:$KALDI_ROOT/src/fstbin/:$KALDI_ROOT/src/gmmbin/:$KALDI_ROOT/src/featbin/:$KALDI_ROOT/src/lm/:$KALDI_ROOT/src/sgmmbin/:$KALDI_ROOT/src/sgmm2bin/:$KALDI_ROOT/src/fgmmbin/:$KALDI_ROOT/src/latbin/:$KALDI_ROOT/src/nnetbin:$KALDI_ROOT/src/nnet2bin/:$KALDI_ROOT/src/online2bin/:$KALDI_ROOT/src/ivectorbin/:$KALDI_ROOT/src/kwsbin:$KALDI_ROOT/src/nnet3bin:$KALDI_ROOT/src/chainbin:$KALDI_ROOT/src/rnnlmbin:$PWD:$PATH
+export LC_ALL=C
+EOF
+ln -sf /opt/kaldi/egs/wsj/s5/steps
+ln -sf /opt/kaldi/egs/wsj/s5/utils
+
+mkdir -p conf data/{lang/phones,finetune} exp extractor
+cp $model/conf/{mfcc,mfcc_hires}.conf conf/
+cp $model/conf/online_cmvn.conf conf/  # Only needed if/for finetune_ivector_extractor
+cp $model/conf/online_cmvn.conf extractor/
+# cp $model/ivector_extractor/final.{ie,dubm,mat} extractor/  # Careful not to overwrite finetuned ivector_extractor!
+cp $model/ivector_extractor/global_cmvn.stats extractor/
+cp $model/conf/online_cmvn_iextractor extractor/ 2>/dev/null || true
+cp $model/conf/splice.conf extractor/splice_opts
+echo "18" > data/lang/oov.int
+cp $model/{words,phones}.txt data/lang/
+cp $model/disambig.int data/lang/phones/
+cp $model/wdisambig_{words,phones}.int data/lang/phones/  # Only needed if/for mkgraph.sh
+echo "3" > $model/frame_subsampling_factor
+
+echo "1:2:3:4:5:6:7:8:9:10:11:12:13:14:15" > data/lang/phones/context_indep.csl
+echo "1:2:3:4:5:6:7:8:9:10:11:12:13:14:15" > data/lang/phones/silence.csl
+
+. path.sh
+
+# ln -sfT $model/tree_sp tree_sp
+rm tree_sp 2> /dev/null || true
+mkdir -p tree_sp
+cp $model/phones.txt tree_sp/
+mkdir -p exp/nnet3_chain/finetune/
+cp -r $model/dict data/  # Only needed if/for finetune_tree
+# cp $model/tree_stuff/topo data/lang/  # Only needed if/for finetune_tree
+# cp $model/tree_stuff/sets.int data/lang/phones/  # Only needed if/for finetune_tree
+
+# Skip train.py::create_phone_lm()
+touch tree_sp/ali.1.gz tree_sp/tree tree_sp/final.mdl  # Fake empty, to pacify the training script later
+
+# Skip train.py::create_denominator_fst()
+copy-transition-model $model/final.mdl exp/nnet3_chain/finetune/0.trans_mdl 2> /dev/null
+cp $model/tree $model/tree_stuff/{den,normalization}.fst exp/nnet3_chain/finetune/
+
+perl -ane '@A=split(" ",$_); $w = shift @A; $p = shift @A; @A>0||die;
+    if(@A==1) { print "$w $p $A[0]_S\n"; } else { print "$w $p $A[0]_B ";
+    for($n=1;$n<@A-1;$n++) { print "$A[$n]_I "; } print "$A[$n]_E\n"; } ' \
+    < $model/lexiconp.txt > data/lang/lexiconp_pdp.txt || exit 1;
+utils/lang/make_lexicon_fst.py --sil-prob=0.5 --sil-phone=SIL data/lang/lexiconp_pdp.txt | \
+    fstcompile --isymbols=$model/phones.txt --osymbols=$model/words.txt --keep_isymbols=false --keep_osymbols=false | \
+    fstarcsort --sort_type=olabel > data/lang/L.fst || exit 1
+
+cp -r $dataset/{text,wav.scp,utt2spk} data/finetune
+# ln -sfT /mnt/input/audio_data audio_data
+
+# utils/fix_data_dir.sh data/finetune
+$nice_cmd bash run_finetune_tdnn_1a_daanzu.sh --src-dir $model --extractor-dir extractor --tree-dir tree_sp --nj $(nproc) $*
+
+# > cp -r work.test.per/data/lang/phones/* work.test.fin/data/lang/phones/
+# > cp -r work.test.per/data/lang_chain/topo work.test.fin/data/lang/
diff --git a/run.personal.sh b/run.personal.sh
new file mode 100644
index 0000000..9cfcba9
--- /dev/null
+++ b/run.personal.sh
@@ -0,0 +1,68 @@
+# docker run -it --rm -v $(pwd):/mnt/input -v $(pwd)/work:/mnt/work -w /mnt/work --user "$(id -u):$(id -g)" daanzu/kaldi_ag_training:2020-11-28 bash run.personal.sh models/kaldi_model_daanzu_20200905_1ep-mediumlm data/standard2train --num-epochs 5 --stage -10
+# docker run -it --rm -v $(pwd):/mnt/input -v $(pwd)/work:/mnt/work -w /mnt/work --user "$(id -u):$(id -g)" --runtime=nvidia daanzu/kaldi_ag_training_gpu:2020-11-28 bash run.personal.sh models/kaldi_model_daanzu_20200905_1ep-mediumlm data/standard2train --num-epochs 5 --stage -10
+
+set -euxo pipefail
+
+nice_cmd="nice ionice -c idle"
+stage=-10
+gmm_stage=0  # always stage+10
+
+# Scan through arguments, checking for stage argument, which if included we need to use to set the gmm_stage
+POSITIONAL=()
+while [[ $# -gt 0 ]]
+do
+key="$1"
+case $key in
+    --stage)
+    stage="$2"
+    gmm_stage=$((stage+10))
+    POSITIONAL+=("$1" "$2") # save it in an array for later
+    shift # past argument
+    shift # past value
+    ;;
+    *)    # unknown option
+    POSITIONAL+=("$1") # save it in an array for later
+    shift # past argument
+    ;;
+esac
+done
+set -- "${POSITIONAL[@]}" # restore positional parameters
+
+[[ $# -ge 2 ]] || exit 1
+
+model=/mnt/input/$1; shift
+dataset=/mnt/input/$1; shift
+
+[[ -d $model ]] || exit 1
+[[ -d $dataset ]] || exit 1
+
+echo "base_model=${model#/mnt/input/}" >> params.txt
+echo "train_dataset=${dataset#/mnt/input/}" >> params.txt
+
+cat <<\EOF > cmd.sh
+export train_cmd="utils/run.pl"
+export decode_cmd="utils/run.pl"
+export cuda_cmd="utils/run.pl"
+# export cuda_cmd="utils/run.pl -l gpu=1"
+EOF
+cat <<\EOF > path.sh
+export KALDI_ROOT=/opt/kaldi
+export LD_LIBRARY_PATH="$KALDI_ROOT/tools/openfst/lib:$KALDI_ROOT/tools/openfst/lib/fst:$KALDI_ROOT/src/lib:$LD_LIBRARY_PATH"
+export PATH=$KALDI_ROOT/src/lmbin/:$KALDI_ROOT/../kaldi_lm/:$PWD/utils/:$KALDI_ROOT/src/bin:$KALDI_ROOT/tools/openfst/bin:$KALDI_ROOT/src/fstbin/:$KALDI_ROOT/src/gmmbin/:$KALDI_ROOT/src/featbin/:$KALDI_ROOT/src/lm/:$KALDI_ROOT/src/sgmmbin/:$KALDI_ROOT/src/sgmm2bin/:$KALDI_ROOT/src/fgmmbin/:$KALDI_ROOT/src/latbin/:$KALDI_ROOT/src/nnetbin:$KALDI_ROOT/src/nnet2bin/:$KALDI_ROOT/src/online2bin/:$KALDI_ROOT/src/ivectorbin/:$KALDI_ROOT/src/kwsbin:$KALDI_ROOT/src/nnet3bin:$KALDI_ROOT/src/chainbin:$KALDI_ROOT/src/rnnlmbin:$PWD:$PATH
+export LC_ALL=C
+EOF
+ln -sf /opt/kaldi/egs/wsj/s5/steps
+ln -sf /opt/kaldi/egs/wsj/s5/utils
+
+mkdir -p data/train data/dict conf exp
+cp $model/conf/{mfcc,mfcc_hires,online_cmvn}.conf conf
+cp $model/dict/{extra_questions.txt,lexiconp.txt,lexicon.txt,nonsilence_phones.txt,optional_silence.txt,silence_phones.txt} data/dict
+
+[[ $stage -gt -10 ]] || rm -rf data/train/*
+cp $dataset/{text,wav.scp,utt2spk} data/train
+# ln -sfT /mnt/input/audio_data audio_data
+# ln -sfT /mnt/input/audio_data/daanzu wav
+
+# utils/fix_data_dir.sh data/train
+$nice_cmd bash run_personal_gmm.sh --nj $(nproc) --stage $gmm_stage
+$nice_cmd bash run_personal_chain_tdnn_1h.sh --nj $(nproc) $*
diff --git a/run_finetune_tdnn_1a_daanzu.sh b/run_finetune_tdnn_1a_daanzu.sh
new file mode 100644
index 0000000..3918e5f
--- /dev/null
+++ b/run_finetune_tdnn_1a_daanzu.sh
@@ -0,0 +1,305 @@
+# !/bin/bash
+
+# Adapted from egs/aishell2/s5/local/nnet3/tuning/finetune_tdnn_1a.sh commit 42a673a5e7f201736dfbf2116e8eaa94745e5a5f
+# Also see:
+#   egs/rm/s5/local/chain/tuning/run_tdnn_wsj_rm_1b.sh
+#   egs/rm/s5/local/chain/tuning/run_tdnn_wsj_rm_1c.sh
+
+# This script uses weight transfer as a transfer learning method to transfer already trained neural net model to a finetune dataset.
+
+# Usage: run_finetune_tdnn_1a_daanzu.sh --src-dir export/tdnn_f.1ep --num-epochs 5 --stage 1 --train-stage -10
+
+# Required Inputs:
+#   data/finetune (text wav.scp utt2spk)
+#   src_dir
+#   tree_dir (tree final.mdl ali.*.gz phones.txt)
+#   lang_dir (oov.int L.fst words.txt phones.txt phones/disambig.int)
+#   conf_dir (mfcc.conf mfcc_hires.conf)
+#   extractor_dir (final.ie final.dubm final.mat global_cmvn.stats splice_opts online_cmvn.conf online_cmvn_iextractor?)
+# Writes To:
+#   data/finetune, data/finetune_hires, data/finetune_sp, data/finetune_sp_hires,
+#   exp/make_mfcc_chain/finetune, exp/make_mfcc_chain/finetune_sp_hires, exp/make_mfcc_chain/finetune_hires,
+#   exp/nnet3_chain/ivectors_finetune_hires, exp/finetune_lats, exp/nnet3_chain/finetune
+
+set -e
+
+data_set=finetune
+data_dir=data/${data_set}
+conf_dir=conf
+lang_dir=data/lang  # FIXME: lang_chain?
+extractor_dir=exp/nnet3_chain/extractor
+# ali_dir=exp/${data_set}_ali
+# lat_dir=exp/${data_set}_lats
+src_dir=exp/nnet3_chain/tdnn_f
+tree_dir=exp/nnet3_chain/tree_sp
+# dir=${src_dir}_${data_set}
+dir=exp/nnet3_chain/${data_set}
+
+respect_speaker_info=false
+finetune_ivector_extractor=false
+finetune_phonelm=false
+
+num_gpus=1
+num_epochs=5
+# initial_lrate=0.0005
+# final_lrate=0.00002
+initial_lrate=.00025
+final_lrate=.000025
+minibatch_size=128,64
+primary_lr_factor=0.25  # learning-rate factor for all except last layer in transferred source model (last layer is 1.0)
+
+xent_regularize=0.1
+train_stage=-4  # Normally default -10, but here -4 to skip phone_LM and den.fst generation training stages.
+get_egs_stage=-10
+common_egs_dir=  # you can set this to use previously dumped egs.
+egs_opts="--num-utts-subset 3000 --max-jobs-run 4 --max-shuffle-jobs-run 10"  # --num-utts-subset 3000 --max-jobs-run 4 --max-shuffle-jobs-run 10
+dropout_schedule='0,0@0.20,0.5@0.50,0'
+frames_per_eg=150,110,100
+chain_left_tolerance=1
+chain_right_tolerance=1
+
+stage=1
+nj=8
+
+echo "$0 $@"  # Print the command line for logging
+. ./path.sh
+. ./cmd.sh
+. ./utils/parse_options.sh
+
+if [ "$num_gpus" -eq 0 ]; then
+  gpu_opt="no"
+  num_gpus=1
+else
+  gpu_opt="wait"
+fi
+
+function write_params() {
+  for v in $*; do
+    echo "${v}=${!v}" >> params.txt
+  done
+}
+
+function log_stage() {
+  echo
+  echo "# Stage $1"
+  [ -z "$2" ] && echo "# $2"
+  echo "# $(date)"
+  echo
+}
+
+function compute_features() {
+  # compute_features <conf_file> [input_data_dir_affix]
+  rm -f ${data_dir}${2}/{cmvn.scp,feats.scp}
+  steps/make_mfcc.sh \
+    --cmd "$train_cmd" --nj $nj --mfcc-config ${1} \
+    ${data_dir}${2} exp/make_mfcc_chain/${data_set}${2}.log exp/make_mfcc_chain
+  steps/compute_cmvn_stats.sh ${data_dir}${2} exp/make_mfcc_chain/${data_set}${2}.log exp/make_mfcc_chain || exit 1;
+  utils/fix_data_dir.sh ${data_dir}${2} || exit 1;
+}
+
+if [ $stage -le 1 ]; then
+  log_stage 1 "Compute features (MFCC & CMVN stats) of the new dataset, including perturbing the data"
+  # (Approximately 0.66min single-core compute time per core per 1hr audio data)
+  utils/fix_data_dir.sh ${data_dir} || exit 1;
+
+  # Standard lores training data
+  compute_features $conf_dir/mfcc.conf
+
+  rm -rf ${data_dir}_sp
+  utils/data/perturb_data_dir_speed_3way.sh ${data_dir} ${data_dir}_sp || exit 1;
+  compute_features $conf_dir/mfcc.conf _sp
+
+  rm -rf ${data_dir}_sp_hires
+  utils/copy_data_dir.sh ${data_dir}_sp ${data_dir}_sp_hires
+  utils/data/perturb_data_dir_volume.sh ${data_dir}_sp_hires || exit 1;
+  compute_features $conf_dir/mfcc_hires.conf _sp_hires
+
+  rm -rf ${data_dir}_sp_novp_hires
+  utils/copy_data_dir.sh ${data_dir}_sp ${data_dir}_sp_novp_hires
+  compute_features $conf_dir/mfcc_hires.conf _sp_novp_hires
+
+  rm -rf ${data_dir}_vp_hires
+  utils/copy_data_dir.sh ${data_dir} ${data_dir}_vp_hires
+  utils/data/perturb_data_dir_volume.sh ${data_dir}_vp_hires || exit 1;
+  compute_features $conf_dir/mfcc_hires.conf _vp_hires
+
+  rm -rf ${data_dir}_hires
+  utils/copy_data_dir.sh ${data_dir} ${data_dir}_hires
+  compute_features $conf_dir/mfcc_hires.conf _hires
+fi
+
+train_affix=_hires
+# train_affix=_sp_novp_hires
+train_data_dir=${data_dir}${train_affix}
+train_ivector_dir=exp/nnet3_chain/ivectors_${data_set}${train_affix}
+lat_dir=exp/nnet3_chain/lats_${data_set}${train_affix}
+lores_train_data_dir=${data_dir}_sp
+# extractor_dir=exp/nnet3_chain/extractor_${data_set}${train_affix}
+
+if $finetune_ivector_extractor; then
+  train_set=${data_set}
+
+  if [ $stage -le 2 ]; then
+    log_stage 2 "Finetune ivectors: Train diagonal UBM"
+
+    echo "$0: computing a subset of data to train the diagonal UBM."
+    # We'll use about a quarter of the data.
+    temp_data_root=exp/nnet3_chain/diag_ubm
+    mkdir -p $temp_data_root
+
+    num_utts_total=$(wc -l <data/${train_set}${train_affix}/utt2spk)
+    num_utts=$[$num_utts_total/4]
+    utils/data/subset_data_dir.sh data/${train_set}${train_affix} \
+       $num_utts ${temp_data_root}/${train_set}${train_affix}_subset
+
+    echo "$0: computing a PCA transform from the hires data."
+    steps/online/nnet2/get_pca_transform.sh --cmd "$train_cmd" \
+        --splice-opts "--left-context=3 --right-context=3" \
+        --max-utts 10000 --subsample 2 \
+         ${temp_data_root}/${train_set}${train_affix}_subset \
+         exp/nnet3_chain/pca_transform
+
+    echo "$0: training the diagonal UBM."
+    # Use 512 Gaussians in the UBM.
+    steps/online/nnet2/train_diag_ubm.sh --cmd "$train_cmd" --nj $nj \
+      --num-frames 700000 \
+      --num-threads 8 \
+      ${temp_data_root}/${train_set}${train_affix}_subset 512 \
+      exp/nnet3_chain/pca_transform $temp_data_root
+  fi
+
+  if [ $stage -le 3 ]; then
+    log_stage 3 "Finetune ivectors: Train ivector extractor"
+    # Train the iVector extractor.  Use all of the speed-perturbed data since iVector extractors
+    # can be sensitive to the amount of data.  The script defaults to an iVector dimension of
+    # 100.
+    echo "$0: training the iVector extractor"
+    steps/online/nnet2/train_ivector_extractor.sh --cmd "$train_cmd" --nj $nj \
+       data/${train_set}${train_affix} $temp_data_root \
+       $extractor_dir || exit 1;
+  fi
+
+  if [ $stage -le 4 ]; then
+    log_stage 4 "Finetune ivectors: Extract ivectors of the new dataset with newly-trained extractor"
+    # We extract iVectors on the speed-perturbed training data after combining
+    # short segments, which will be what we train the system on.  With
+    # --utts-per-spk-max 2, the script pairs the utterances into twos, and treats
+    # each of these pairs as one speaker; this gives more diversity in iVectors..
+    # Note that these are extracted 'online'.
+
+    # note, we don't encode the 'max2' in the name of the ivectordir even though
+    # that's the data we extract the ivectors from, as it's still going to be
+    # valid for the non-'max2' data, the utterance list is the same.
+
+    # having a larger number of speakers is helpful for generalization, and to
+    # handle per-utterance decoding well (iVector starts at zero).
+    ivectordir=exp/nnet3_chain/ivectors_${train_set}${train_affix}
+    temp_data_root=${ivectordir}
+    utils/data/modify_speaker_info.sh --utts-per-spk-max 2 --respect-speaker-info $respect_speaker_info \
+      data/${train_set}${train_affix} ${temp_data_root}/${train_set}${train_affix}_max2
+
+    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $nj \
+      ${temp_data_root}/${train_set}${train_affix}_max2 \
+      $extractor_dir $ivectordir
+  fi
+
+else
+  if [ $stage -le 4 ]; then
+    log_stage 4 "Extract ivectors of the new dataset using source model's extractor"
+    # (Approximately 0.066min single-core compute time per core per 1hr audio data)
+    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $nj \
+      $train_data_dir $extractor_dir $train_ivector_dir
+  fi
+fi
+
+if [ $stage -le 5 ]; then
+  log_stage 5 "Align the new dataset with source NN"
+  # (Approximately 0.085hr single-core compute time per core per 1hr audio data)
+
+  # steps/nnet3/align.sh --cmd "$train_cmd" --nj ${nj} ${data_dir} $lang_dir ${src_dir} ${ali_dir}
+
+  steps/nnet3/align_lats.sh --cmd "$train_cmd" --nj $nj \
+    --acoustic-scale 1.0 \
+    --scale-opts '--transition-scale=1.0 --self-loop-scale=1.0' \
+    --online-ivector-dir $train_ivector_dir \
+    --generate_ali_from_lats true \
+    $train_data_dir $lang_dir ${src_dir} ${lat_dir}
+    # --frames-per-chunk 150 \
+  rm $lat_dir/fsts.*.gz # save space
+
+  # FIXME: steps/nnet3/chain/align_lats.sh?
+  # steps/nnet3/chain/align_lats.sh --cmd "$train_cmd" --nj $nj \
+  #   --online-ivector-dir $train_ivector_dir \
+  #   $train_data_dir $lang_dir ${src_dir} ${lat_dir}
+fi
+
+# NOTE: We must use the same tree as was used to train the nnet model (need the same num_pdfs).
+
+if $finetune_phonelm; then
+  if [ $stage -le 6 ]; then
+    log_stage 6 "Copy data to allow building new finetuned phone_lm"
+    # Requires: tree_sp/{tree,final.mdl}
+    cp $lat_dir/ali.*.gz $lat_dir/num_jobs $tree_dir
+    # Note: actually do it by $train_stage <= -6
+  fi
+fi
+
+if [ $stage -le 8 ]; then
+  log_stage 8 "Copy source NN"
+  # Set the learning-rate-factor for all transferred layers but the last output layer to primary_lr_factor.
+  $train_cmd $dir/log/generate_input_mdl.log \
+    nnet3-am-copy --raw=true \
+      --edits="set-learning-rate-factor name=* learning-rate-factor=$primary_lr_factor; set-learning-rate-factor name=output* learning-rate-factor=1.0" \
+      $src_dir/final.mdl $dir/input.raw || exit 1;
+fi
+
+# echo "$0: sleeping..."
+# sleep 600
+
+if [ $stage -le 9 ]; then
+  log_stage 9 "Train new fine-tuned NN"
+
+  # Exclude phone_LM and den.fst generation training stages.
+  # if [ $train_stage -lt -4 ]; then train_stage=-4 ; fi
+
+  # we use chain model from source to generate lats for target and the
+  # tolerance used in chain egs generation using this lats should be 1 or 2 which is
+  # (source_egs_tolerance/frame_subsampling_factor)
+  # source_egs_tolerance = 5
+  # chain_opts=(--chain.alignment-subsampling-factor=1 --chain.left-tolerance=1 --chain.right-tolerance=1)
+  chain_opts=(--chain.alignment-subsampling-factor=1 --chain.left-tolerance=$chain_left_tolerance --chain.right-tolerance=$chain_right_tolerance)
+
+  write_params chunk_width dropout_schedule xent_regularize initial_lrate final_lrate num_epochs num_gpu_jobs stage train_stage num_utts_subset
+
+  steps/nnet3/chain/train.py --stage $train_stage ${chain_opts[@]} \
+    --cmd "$decode_cmd" \
+    --trainer.input-model $dir/input.raw \
+    --feat.online-ivector-dir $train_ivector_dir \
+    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
+    --chain.xent-regularize $xent_regularize \
+    --chain.leaky-hmm-coefficient 0.1 \
+    --chain.l2-regularize 0.0 \
+    --chain.apply-deriv-weights false \
+    --chain.lm-opts="--num-extra-lm-states=2000" \
+    --trainer.dropout-schedule $dropout_schedule \
+    --trainer.add-option="--optimization.memory-compression-level=2" \
+    --egs.dir "$common_egs_dir" \
+    --egs.stage $get_egs_stage \
+    --egs.opts "--frames-overlap-per-eg 0 --constrained false $egs_opts" \
+    --egs.chunk-width $frames_per_eg \
+    --egs.nj $nj \
+    --trainer.num-chunk-per-minibatch $minibatch_size \
+    --trainer.frames-per-iter 1500000 \
+    --trainer.num-epochs $num_epochs \
+    --trainer.optimization.num-jobs-initial $num_gpus \
+    --trainer.optimization.num-jobs-final $num_gpus \
+    --trainer.optimization.initial-effective-lrate $initial_lrate \
+    --trainer.optimization.final-effective-lrate $final_lrate \
+    --trainer.max-param-change 2.0 \
+    --use-gpu $gpu_opt \
+    --cleanup.remove-egs false \
+    --feat-dir $train_data_dir \
+    --tree-dir $tree_dir \
+    --lat-dir $lat_dir \
+    --dir $dir  || exit 1;
+fi
diff --git a/run_personal_chain_tdnn_1h.sh b/run_personal_chain_tdnn_1h.sh
new file mode 100644
index 0000000..5835982
--- /dev/null
+++ b/run_personal_chain_tdnn_1h.sh
@@ -0,0 +1,471 @@
+#!/bin/bash
+
+# Adapted from egs/mini_librispeech/s5/local/chain/tuning/run_tdnn_1h.sh
+
+# Required Inputs:
+#   conf/{mfcc,mfcc_hires,online_cmvn}.conf
+#   data/train/{text,wav.scp,utt2spk}
+#   data/lang/?????
+# Writes To:
+#   exp/?????
+
+# Set -e here so that we catch if any executable fails immediately
+set -euo pipefail
+
+# First the options that are passed through to run_ivector_common.sh
+# (some of which are also used in this script directly).
+stage=0
+nj=8
+train_set=train
+test_sets=test
+gmm=tri3b
+nnet3_affix=
+
+# The rest are configs specific to this script.  Most of the parameters
+# are just hardcoded at this level, in the commands below.
+affix=1h   # affix for the TDNN directory name
+tree_affix=
+train_stage=-10
+get_egs_stage=-10
+decode_iter=
+
+# training options
+chunk_width=140,100,160
+# chunk_width=140,100,160,40
+num_utts_subset=300  # default 300; you may want many more for short-utterance datasets <----------------
+dropout_schedule='0,0@0.20,0.3@0.50,0'
+common_egs_dir=
+xent_regularize=0.1
+
+# training options
+srand=0
+# remove_egs=true
+remove_egs=false
+reporting_email=
+
+#decode options
+test_online_decoding=true  # if true, it will run the last decoding stage.
+
+# daanzu options
+tdnnf_dim=1024
+decode_name=
+respect_speaker_info=true
+initial_lrate=
+final_lrate=
+num_epochs=20
+num_gpus=1
+spot=
+
+
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+num_gpu_jobs=$num_gpus
+
+function write_params() {
+  for v in $*; do
+    echo "${v}=${!v}" >> params.txt
+  done
+}
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+if [ $stage -le 0 ]; then
+  # Clear for run_personal_nnet3_ivector_common.sh below
+  rm -rf data/train_sp data/train_sp_hires
+fi
+
+# The iVector-extraction and feature-dumping parts are the same as the standard
+# nnet3 setup, and you can skip them by setting "--stage 11" if you have already
+# run those things.
+bash run_personal_nnet3_ivector_common.sh --stage $stage \
+                                  --nj $nj \
+                                  --train-set $train_set \
+                                  --gmm $gmm \
+                                  --respect-speaker-info $respect_speaker_info \
+                                  --nnet3-affix "$nnet3_affix" || exit 1;
+
+# Problem: We have removed the "train_" prefix of our training set in
+# the alignment directory names! Bad!
+gmm_dir=exp/$gmm
+ali_dir=exp/${gmm}_ali_${train_set}_sp
+tree_dir=exp/chain${nnet3_affix}/tree_sp${tree_affix:+_$tree_affix}
+lang=data/lang_chain
+lat_dir=exp/chain${nnet3_affix}/${gmm}_${train_set}_sp_lats
+dir=exp/chain${nnet3_affix}/tdnn${affix}_sp
+train_data_dir=data/${train_set}_sp_hires
+lores_train_data_dir=data/${train_set}_sp
+train_ivector_dir=exp/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires
+
+for f in $gmm_dir/final.mdl $train_data_dir/feats.scp $train_ivector_dir/ivector_online.scp \
+    $lores_train_data_dir/feats.scp $ali_dir/ali.1.gz; do
+  [ ! -f $f ] && echo "$0: expected file $f to exist" && exit 1
+done
+
+if [ $stage -le 10 ]; then
+  echo "$0: creating lang directory $lang with chain-type topology"
+  # Create a version of the lang/ directory that has one state per phone in the
+  # topo file. [note, it really has two states.. the first one is only repeated
+  # once, the second one has zero or more repeats.]
+  rm -rf $lang
+  if [ -d $lang ]; then
+    if [ $lang/L.fst -nt data/lang/L.fst ]; then
+      echo "$0: $lang already exists, not overwriting it; continuing"
+    else
+      echo "$0: $lang already exists and seems to be older than data/lang..."
+      echo " ... not sure what to do.  Exiting."
+      exit 1;
+    fi
+  else
+    cp -r data/lang $lang
+    silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
+    nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
+    # Use our special topology... note that later on may have to tune this
+    # topology.
+    steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
+  fi
+fi
+
+if [ $stage -le 11 ]; then
+  # Get the alignments as lattices (gives the chain training more freedom).
+  # use the same num-jobs as the alignments
+  steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" ${lores_train_data_dir} \
+    data/lang $gmm_dir $lat_dir
+  rm $lat_dir/fsts.*.gz # save space
+fi
+
+if [ $stage -le 12 ]; then
+  # Build a tree using our new topology.  We know we have alignments for the
+  # speed-perturbed data (local/nnet3/run_ivector_common.sh made them), so use
+  # those.  The num-leaves is always somewhat less than the num-leaves from
+  # the GMM baseline.
+  # if [ -f $tree_dir/final.mdl ]; then
+  #    echo "$0: $tree_dir/final.mdl already exists, refusing to overwrite it."
+  #    exit 1;
+  # fi
+  steps/nnet3/chain/build_tree.sh \
+    --frame-subsampling-factor 3 \
+    --context-opts "--context-width=2 --central-position=1" \
+    --cmd "$train_cmd" 3500 ${lores_train_data_dir} \
+    $lang $ali_dir $tree_dir
+fi
+
+if [ $tdnnf_dim -eq 738 ]; then
+  if [ $stage -le 13 ]; then
+    # from mini_librispeech tdnn_1h
+    mkdir -p $dir
+    echo "$0: creating neural net configs using the xconfig parser";
+
+    num_targets=$(tree-info $tree_dir/tree |grep num-pdfs|awk '{print $2}')
+    learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+
+    tdnn_opts="l2-regularize=0.03 dropout-proportion=0.0 dropout-per-dim-continuous=true"
+    tdnnf_opts="l2-regularize=0.03 dropout-proportion=0.0 bypass-scale=0.66"
+    linear_opts="l2-regularize=0.03 orthonormal-constraint=-1.0"
+    prefinal_opts="l2-regularize=0.03"
+    output_opts="l2-regularize=0.015"
+
+    write_params num_targets learning_rate_factor tdnn_opts tdnnf_opts linear_opts prefinal_opts output_opts
+
+    mkdir -p $dir/configs
+    cat <<EOF > $dir/configs/network.xconfig
+    input dim=100 name=ivector
+    input dim=40 name=input
+
+    delta-layer name=delta
+    no-op-component name=input2 input=Append(delta, Scale(1.0, ReplaceIndex(ivector, t, 0)))
+
+    # the first splicing is moved before the lda layer, so no splicing here
+    relu-batchnorm-dropout-layer name=tdnn1 $tdnn_opts dim=768
+    tdnnf-layer name=tdnnf2 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=1
+    tdnnf-layer name=tdnnf3 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=1
+    tdnnf-layer name=tdnnf4 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=1
+    tdnnf-layer name=tdnnf5 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=0
+    tdnnf-layer name=tdnnf6 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=3
+    tdnnf-layer name=tdnnf7 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=3
+    tdnnf-layer name=tdnnf8 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=3
+    tdnnf-layer name=tdnnf9 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=3
+    tdnnf-layer name=tdnnf10 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=3
+    tdnnf-layer name=tdnnf11 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=3
+    tdnnf-layer name=tdnnf12 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=3
+    tdnnf-layer name=tdnnf13 $tdnnf_opts dim=768 bottleneck-dim=96 time-stride=3
+    linear-component name=prefinal-l dim=192 $linear_opts
+
+    ## adding the layers for chain branch
+    prefinal-layer name=prefinal-chain input=prefinal-l $prefinal_opts small-dim=192 big-dim=768
+    output-layer name=output include-log-softmax=false dim=$num_targets $output_opts
+
+    # adding the layers for xent branch
+    prefinal-layer name=prefinal-xent input=prefinal-l $prefinal_opts small-dim=192 big-dim=768
+    output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor $output_opts
+EOF
+    steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+  fi
+  initial_lrate=${initial_lrate:-0.002}
+  final_lrate=${final_lrate:-0.0002}
+
+elif [ $tdnnf_dim -eq 1024 ]; then
+  if [ $stage -le 13 ]; then
+    # from wsj tdnn_1g
+    mkdir -p $dir
+    echo "$0: creating neural net configs using the xconfig parser";
+
+    num_targets=$(tree-info $tree_dir/tree |grep num-pdfs|awk '{print $2}')
+    learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+
+    tdnn_opts="l2-regularize=0.01 dropout-proportion=0.0 dropout-per-dim-continuous=true"
+    tdnnf_opts="l2-regularize=0.01 dropout-proportion=0.0 bypass-scale=0.66"
+    linear_opts="l2-regularize=0.01 orthonormal-constraint=-1.0"
+    prefinal_opts="l2-regularize=0.01"
+    output_opts="l2-regularize=0.005"
+
+    write_params num_targets learning_rate_factor tdnn_opts tdnnf_opts linear_opts prefinal_opts output_opts
+
+    mkdir -p $dir/configs
+    cat <<EOF > $dir/configs/network.xconfig
+    input dim=100 name=ivector
+    input dim=40 name=input
+
+    delta-layer name=delta
+    no-op-component name=input2 input=Append(delta, Scale(1.0, ReplaceIndex(ivector, t, 0)))
+    # no-op-component name=input2 input=delta
+
+    relu-batchnorm-dropout-layer name=tdnn1 $tdnn_opts dim=1024 input=input2
+    tdnnf-layer name=tdnnf2 $tdnnf_opts dim=1024 bottleneck-dim=96 time-stride=1
+    tdnnf-layer name=tdnnf3 $tdnnf_opts dim=1024 bottleneck-dim=96 time-stride=1
+    tdnnf-layer name=tdnnf4 $tdnnf_opts dim=1024 bottleneck-dim=96 time-stride=1
+    tdnnf-layer name=tdnnf5 $tdnnf_opts dim=1024 bottleneck-dim=96 time-stride=0
+    tdnnf-layer name=tdnnf6 $tdnnf_opts dim=1024 bottleneck-dim=96 time-stride=3
+    tdnnf-layer name=tdnnf7 $tdnnf_opts dim=1024 bottleneck-dim=96 time-stride=3
+    tdnnf-layer name=tdnnf8 $tdnnf_opts dim=1024 bottleneck-dim=96 time-stride=3
+    tdnnf-layer name=tdnnf9 $tdnnf_opts dim=1024 bottleneck-dim=96 time-stride=3
+    tdnnf-layer name=tdnnf10 $tdnnf_opts dim=1024 bottleneck-dim=96 time-stride=3
+    tdnnf-layer name=tdnnf11 $tdnnf_opts dim=1024 bottleneck-dim=96 time-stride=3
+    tdnnf-layer name=tdnnf12 $tdnnf_opts dim=1024 bottleneck-dim=96 time-stride=3
+    tdnnf-layer name=tdnnf13 $tdnnf_opts dim=1024 bottleneck-dim=96 time-stride=3
+    linear-component name=prefinal-l dim=192 $linear_opts
+
+    ## adding the layers for chain branch
+    prefinal-layer name=prefinal-chain input=prefinal-l $prefinal_opts small-dim=192 big-dim=1024
+    output-layer name=output include-log-softmax=false dim=$num_targets $output_opts
+
+    # adding the layers for xent branch
+    prefinal-layer name=prefinal-xent input=prefinal-l $prefinal_opts small-dim=192 big-dim=1024
+    output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor $output_opts
+EOF
+    steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+  fi
+  initial_lrate=${initial_lrate:-0.001}
+  final_lrate=${final_lrate:-0.0001}
+  # initial_lrate=0.001
+  # final_lrate=0.00005
+
+elif [ $tdnnf_dim -eq 1536 ]; then
+  if [ $stage -le 13 ]; then
+    # from chime5 tdnn_1b
+    mkdir -p $dir
+    echo "$0: creating neural net configs using the xconfig parser";
+
+    num_targets=$(tree-info $tree_dir/tree |grep num-pdfs|awk '{print $2}')
+    learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+
+    tdnn_opts="l2-regularize=0.01 dropout-proportion=0.0 dropout-per-dim=true dropout-per-dim-continuous=true"
+    tdnnf_opts="l2-regularize=0.01 dropout-proportion=0.0 bypass-scale=0.66"
+    linear_opts="l2-regularize=0.01 orthonormal-constraint=-1.0"
+    prefinal_opts="l2-regularize=0.01"
+    output_opts="l2-regularize=0.002"
+
+    write_params num_targets learning_rate_factor tdnn_opts tdnnf_opts linear_opts prefinal_opts output_opts
+
+    mkdir -p $dir/configs
+    cat <<EOF > $dir/configs/network.xconfig
+    input dim=100 name=ivector
+    input dim=40 name=input
+
+    delta-layer name=delta
+    no-op-component name=input2 input=Append(delta, Scale(1.0, ReplaceIndex(ivector, t, 0)))
+
+    relu-batchnorm-dropout-layer name=tdnn1 $tdnn_opts dim=1536 input=input2
+    tdnnf-layer name=tdnnf2 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=1
+    tdnnf-layer name=tdnnf3 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=1
+    tdnnf-layer name=tdnnf4 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=1
+    tdnnf-layer name=tdnnf5 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=0
+    tdnnf-layer name=tdnnf6 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+    tdnnf-layer name=tdnnf7 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+    tdnnf-layer name=tdnnf8 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+    tdnnf-layer name=tdnnf9 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+    tdnnf-layer name=tdnnf10 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+    tdnnf-layer name=tdnnf11 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+    tdnnf-layer name=tdnnf12 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+    tdnnf-layer name=tdnnf13 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+    tdnnf-layer name=tdnnf14 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+    tdnnf-layer name=tdnnf15 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
+    linear-component name=prefinal-l dim=256 $linear_opts
+
+    prefinal-layer name=prefinal-chain input=prefinal-l $prefinal_opts big-dim=1536 small-dim=256
+    output-layer name=output include-log-softmax=false dim=$num_targets $output_opts
+
+    prefinal-layer name=prefinal-xent input=prefinal-l $prefinal_opts big-dim=1536 small-dim=256
+    output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor $output_opts
+EOF
+    steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+  fi
+  initial_lrate=${initial_lrate:-0.001}
+  final_lrate=${final_lrate:-0.0001}
+
+else
+  echo "$0: ERROR: invalid tdnnf_dim: ${tdnnf_dim}"; exit 1
+fi
+
+
+if [ $stage -le 14 ]; then
+  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
+    utils/create_split_dir.pl \
+     /export/b0{3,4,5,6}/$USER/kaldi-data/egs/mini_librispeech-$(date +'%m_%d_%H_%M')/s5/$dir/egs/storage $dir/egs/storage
+  fi
+  [ $num_gpu_jobs -gt 1 ] && sudo nvidia-smi -c 3
+  write_params chunk_width dropout_schedule xent_regularize initial_lrate final_lrate num_epochs num_gpu_jobs stage train_stage num_utts_subset
+  # --num-valid-egs-combine --num-train-egs-combine --num-egs-diagnostic ??? see steps/nnet3/chain/get_egs.sh
+
+  steps/nnet3/chain/train.py --stage=$train_stage \
+    --cmd="$decode_cmd" \
+    --feat.online-ivector-dir=$train_ivector_dir \
+    --feat.cmvn-opts="--norm-means=false --norm-vars=false" \
+    --chain.xent-regularize $xent_regularize \
+    --chain.leaky-hmm-coefficient=0.1 \
+    --chain.l2-regularize=0.0 \
+    --chain.apply-deriv-weights=false \
+    --chain.lm-opts="--num-extra-lm-states=2000" \
+    --trainer.dropout-schedule $dropout_schedule \
+    --trainer.add-option="--optimization.memory-compression-level=2" \
+    --trainer.srand=$srand \
+    --trainer.max-param-change=2.0 \
+    --trainer.num-epochs=$num_epochs \
+    --trainer.frames-per-iter=3000000 \
+    --trainer.optimization.num-jobs-initial=$num_gpu_jobs \
+    --trainer.optimization.num-jobs-final=$num_gpu_jobs \
+    --trainer.optimization.initial-effective-lrate=$initial_lrate \
+    --trainer.optimization.final-effective-lrate=$final_lrate \
+    --trainer.num-chunk-per-minibatch=128,64 \
+    --egs.chunk-width=$chunk_width \
+    --egs.dir="$common_egs_dir" \
+    --egs.opts="--frames-overlap-per-eg 0 --num-utts-subset $num_utts_subset" \
+    --cleanup.remove-egs=$remove_egs \
+    --cleanup.preserve-model-interval=1000 \
+    --use-gpu=true \
+    --reporting.email="$reporting_email" \
+    --feat-dir=$train_data_dir \
+    --tree-dir=$tree_dir \
+    --lat-dir=$lat_dir \
+    --dir=$dir  || exit 1;
+fi
+
+steps/online/nnet3/prepare_online_decoding.sh \
+  --mfcc-config conf/mfcc_hires.conf \
+  $lang exp/nnet3${nnet3_affix}/extractor ${dir} ${dir}_online
+
+exit 0;
+
+if [ $stage -le 15 ]; then
+  # Note: it's not important to give mkgraph.sh the lang directory with the
+  # matched topology (since it gets the topology file from the model).
+  utils/mkgraph.sh \
+    --self-loop-scale 1.0 data/lang${decode_name:+_$decode_name} \
+    $tree_dir $tree_dir/graph_${decode_name} || exit 1;
+fi
+
+if [ $stage -le 16 ]; then
+  frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
+  rm $dir/.error 2>/dev/null || true
+
+  for data in $test_sets; do
+    (
+      nspk=$(wc -l <data/${data}_hires/spk2utt)
+      steps/nnet3/decode.sh \
+          --acwt 1.0 --post-decode-acwt 10.0 \
+          --frames-per-chunk $frames_per_chunk \
+          --nj $nspk --cmd "$decode_cmd"  --num-threads 4 \
+          --online-ivector-dir exp/nnet3${nnet3_affix}/ivectors_${data}_hires \
+          $tree_dir/graph_${decode_name} data/${data}_hires ${dir}/decode_${decode_name}_${data} || exit 1
+      # steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+      #   data/lang_test_{tgsmall,tglarge} \
+      #  data/${data}_hires ${dir}/decode_{tgsmall,tglarge}_${data} || exit 1
+    ) || touch $dir/.error &
+  done
+  wait
+  [ -f $dir/.error ] && echo "$0: there was a problem while decoding" && exit 1
+fi
+
+# Not testing the 'looped' decoding separately, because for
+# TDNN systems it would give exactly the same results as the
+# normal decoding.
+
+if $test_online_decoding && [ $stage -le 17 ]; then
+  # note: if the features change (e.g. you add pitch features), you will have to
+  # change the options of the following command line.
+  steps/online/nnet3/prepare_online_decoding.sh \
+    --mfcc-config conf/mfcc_hires.conf \
+    $lang exp/nnet3${nnet3_affix}/extractor ${dir} ${dir}_online
+
+  rm $dir/.error 2>/dev/null || true
+
+  for data in $test_sets; do
+    (
+      nspk=$(wc -l <data/${data}_hires/spk2utt)
+      # note: we just give it "data/${data}" as it only uses the wav.scp, the
+      # feature type does not matter.
+      steps/online/nnet3/decode.sh \
+        --acwt 1.0 --post-decode-acwt 10.0 \
+        --nj $nspk --cmd "$decode_cmd" \
+        $tree_dir/graph_${decode_name} data/${data} ${dir}_online/decode_${decode_name}_${data} || exit 1
+      # steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+      #   data/lang_test_{tgsmall,tglarge} \
+      #  data/${data}_hires ${dir}_online/decode_{tgsmall,tglarge}_${data} || exit 1
+    ) || touch $dir/.error &
+  done
+  wait
+  [ -f $dir/.error ] && echo "$0: there was a problem while decoding" && exit 1
+fi
+
+if [ $stage -le 18 ]; then
+  export_dir=export/$(basename ${dir}_online)
+  [ -d $export_dir ] && echo "$0: backing up $export_dir" && mv $export_dir ${export_dir}.$(date +"%Y%m%d_%H%M%S")
+  echo "$0: exporting to $export_dir";
+  mkdir -p $export_dir
+  files="conf/  final.mdl  frame_subsampling_factor  ivector_extractor/  phones.txt  tree"
+  for file in $files; do
+    cp -rp ${dir}_online/$file $export_dir
+  done
+  cp -rp data/lang${decode_name:+_$decode_name} $export_dir
+  mkdir -p $export_dir/local
+  cp -rp data/local/dict $export_dir/local
+  cp -rp data/local/lang $export_dir/local
+  cp -rp $tree_dir/graph_${decode_name} $export_dir/graph
+  # cp -rp $tree_dir/1.mdl $export_dir
+  cp -rp local/chain_run_tdnn_1h.sh $export_dir
+  cp -rp $dir/accuracy.report $export_dir
+  cp -rp $0 $export_dir
+fi
+
+exit 0;
+
+
+export affix=1h
+export nnet3_affix=.1h
+export dir=exp/chain${nnet3_affix}/tdnn${affix}_sp
+export tree_dir=exp/chain${nnet3_affix}/tree_sp${tree_affix:+_$tree_affix}
+export data=test
+steps/online/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 --nj 64 --cmd "run.pl" $tree_dir/graph_${decode_name} data/${data} ${dir}_online/decode_${decode_name}_${data}
+steps/online/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 --nj 64 --cmd "run.pl" --iter ../tdnn1h_sp/60-final $tree_dir/graph_${decode_name} data/${data} ${dir}_online/decode_${decode_name}_${data}
+for x in exp/*/decode*; do [ -d $x ] && [[ $x =~ "$1" ]] && grep WER $x/wer_* | utils/best_wer.sh; done
+for x in exp/chain/*/decode*; do [ -d $x ] && [[ $x =~ "$1" ]] && grep WER $x/wer_* | utils/best_wer.sh; done
diff --git a/run_personal_gmm.sh b/run_personal_gmm.sh
new file mode 100644
index 0000000..e8f0ef2
--- /dev/null
+++ b/run_personal_gmm.sh
@@ -0,0 +1,104 @@
+#!/bin/bash
+
+# Adapted from egs/rm/s5/run.sh
+
+# Usage: run_personal_start.sh arpa_file openwebtext.arpa.gz --stage 0
+
+# Required Inputs:
+#   conf/mfcc.conf
+#   data/train/{text,wav.scp,utt2spk}
+#   dict/{extra_questions.txt,lexiconp.txt,lexicon.txt,nonsilence_phones.txt,optional_silence.txt,silence_phones.txt}
+# Writes To:
+#   data/local/dict/
+#   data/lang/
+#   exp/{make_mfcc,mono,mono_ali,tri1,tri1_ali,tri2b,tri2b_ali,tri3b}
+
+set -ev
+
+stage=0
+endstage=99
+nj=16
+test_decoding=false
+dataset=train
+dict_dir=data/dict
+arpa_file=
+. utils/parse_options.sh  # e.g. this parses the --stage option if supplied.
+
+. ./path.sh
+. ./cmd.sh
+
+if [ $stage -le 0 ] && [ $endstage -ge 0 ]; then
+    # data preparation.
+    rm -f data/train/{spk2utt,feats.scp,cmvn.scp}
+    rm -rf data/local data/lang
+
+    utils/fix_data_dir.sh data/$dataset || exit 1;
+    # utils/utt2spk_to_spk2utt.pl data/$dataset/utt2spk > data/$dataset/spk2utt
+    featdir=mfcc
+    steps/make_mfcc.sh --nj $nj --cmd "$train_cmd" data/$dataset exp/make_mfcc/$dataset $featdir
+    # steps/make_mfcc.sh --nj $nj --cmd "$train_cmd" --allow-downsample true data/$dataset exp/make_mfcc/$dataset $featdir
+    steps/compute_cmvn_stats.sh data/$dataset exp/make_mfcc/$dataset $featdir
+    utils/validate_data_dir.sh data/$dataset
+
+    mkdir -p data/local/dict
+    cp $dict_dir/* data/local/dict/
+    utils/prepare_lang.sh data/local/dict "<unk>" data/local/lang data/lang
+    utils/validate_lang.pl data/lang
+
+    # if [ -z "$arpa_file" ]; then
+    #     mkdir data/local/tmp
+    #     ngram-count -order 3 -write-vocab data/local/tmp/vocab-full.txt -wbdiscount -text data/local/dict/corpus.txt -lm data/local/tmp/lm.arpa
+    #     arpa2fst --disambig-symbol=#0 --read-symbol-table=data/lang/words.txt data/local/tmp/lm.arpa data/lang/G.fst
+    #     # ../kenlm/lmplz --text dict/corpus.txt --arpa data/local/tmp/lm.arpa -S 50% -o 3
+    # else
+    #     zcat -f "$arpa_file" | arpa2fst --disambig-symbol=#0 --read-symbol-table=data/lang/words.txt - data/lang/G.fst
+    # fi
+fi
+
+if [ $stage -le 1 ] && [ $endstage -ge 1 ]; then
+    # monophone
+    steps/train_mono.sh --nj $nj --cmd "$train_cmd" data/train data/lang exp/mono
+    # utils/mkgraph.sh data/lang exp/mono exp/mono/graph
+fi
+
+if [ $stage -le 2 ] && [ $endstage -ge 2 ]; then
+    # tri1 [first triphone pass]
+    steps/align_si.sh --nj $nj --cmd "$train_cmd" data/train data/lang exp/mono exp/mono_ali
+    steps/train_deltas.sh --cmd "$train_cmd" 1800 9000 data/train data/lang exp/mono_ali exp/tri1
+    # utils/mkgraph.sh data/lang exp/tri1 exp/tri1/graph
+fi
+
+if [ $stage -le 3 ] && [ $endstage -ge 3 ]; then
+    # tri2b [LDA+MLLT] aka "tri3"
+    steps/align_si.sh --nj $nj --cmd "$train_cmd" --use-graphs true data/train data/lang exp/tri1 exp/tri1_ali
+    steps/train_lda_mllt.sh --cmd "$train_cmd" --splice-opts "--left-context=3 --right-context=3" 1800 9000 data/train data/lang exp/tri1_ali exp/tri2b
+    # utils/mkgraph.sh data/lang exp/tri2b exp/tri2b/graph
+fi
+
+if [ $stage -le 4 ] && [ $endstage -ge 4 ]; then
+    # tri3b [LDA+MLLT+SAT] aka "tri4"?
+    steps/align_si.sh --nj $nj --cmd "$train_cmd" --use-graphs true data/train data/lang exp/tri2b exp/tri2b_ali
+    #????? steps/align_fmllr.sh --nj 8 --cmd "$train_cmd" --use-graphs true data/train data/lang exp/tri2b exp/tri2b_ali
+    steps/train_sat.sh 1800 9000 data/train data/lang exp/tri2b_ali exp/tri3b
+    # utils/mkgraph.sh data/lang exp/tri3b exp/tri3b/graph
+    # utils/mkgraph.sh data/lang_ug exp/tri3b exp/tri3b/graph_ug
+    # steps/decode_fmllr.sh --config conf/decode.config --nj 1 --num-threads 8 --cmd "$decode_cmd" exp/tri3b/graph_ug data/test exp/tri3b/decode_ug
+    # steps/cleanup/find_bad_utts.sh --nj 1 --cmd "$train_cmd" data/train data/lang exp/tri3b_ali exp/tri3b_cleanup
+    # head exp/tri3b_cleanup/all_info.sorted.txt
+fi
+
+# if [ $stage -le 5 ] && [ $endstage -ge 5 ]; then
+#     # tri3b_mmi [LDA+MLLT+SAT+MMI] aka "tri4_mmi"
+#     steps/align_fmllr.sh --nj $nj --cmd "$train_cmd" --use-graphs true data/train data/lang exp/tri3b exp/tri3b_ali
+#     steps/make_denlats.sh --config conf/decode.config --nj $nj --cmd "$train_cmd" --transform-dir exp/tri3b_ali data/train data/lang exp/tri3b exp/tri3b_denlats
+#     steps/train_mmi.sh data/train data/lang exp/tri3b_ali exp/tri3b_denlats exp/tri3b_mmi
+# fi
+
+# if [ $stage -le 8 ] && [ $endstage -ge 8 ]; then
+#     local/kaldi/run_personal_chain_tdnn_1h.sh --stage 0 --num-epochs 20
+# fi
+
+exit
+
+# for x in exp/*/decode*; do [ -d $x ] && [[ $x =~ "$1" ]] && grep WER $x/wer_* | utils/best_wer.sh; done
+# for x in exp/chain/*/decode*; do [ -d $x ] && [[ $x =~ "$1" ]] && grep WER $x/wer_* | utils/best_wer.sh; done
diff --git a/run_personal_nnet3_ivector_common.sh b/run_personal_nnet3_ivector_common.sh
new file mode 100644
index 0000000..194b5c0
--- /dev/null
+++ b/run_personal_nnet3_ivector_common.sh
@@ -0,0 +1,153 @@
+#!/bin/bash
+
+# Adapted from egs/mini_librispeech/s5/local/nnet3/run_ivector_common.sh
+
+set -euo pipefail
+
+# This script is called from local/nnet3/run_tdnn.sh and
+# local/chain/run_tdnn.sh (and may eventually be called by more
+# scripts).  It contains the common feature preparation and
+# iVector-related parts of the script.  See those scripts for examples
+# of usage.
+
+stage=0
+nj=8
+train_set=train
+test_sets=
+gmm=tri3b
+
+nnet3_affix=
+respect_speaker_info=true  # You may want false for one-to-one utt2spk.
+
+. ./cmd.sh
+. ./path.sh
+. utils/parse_options.sh
+
+gmm_dir=exp/${gmm}
+ali_dir=exp/${gmm}_ali_${train_set}_sp
+# ali_dir=exp/${gmm}_ali_sp
+
+for f in data/${train_set}/feats.scp ${gmm_dir}/final.mdl; do
+  if [ ! -f $f ]; then
+    echo "$0: expected file $f to exist"
+    exit 1
+  fi
+done
+
+if [ $stage -le 1 ]; then
+  # Although the nnet will be trained by high resolution data, we still have to
+  # perturb the normal data to get the alignment _sp stands for speed-perturbed
+  echo "$0: preparing directory for low-resolution speed-perturbed data (for alignment)"
+  utils/data/perturb_data_dir_speed_3way.sh data/${train_set} data/${train_set}_sp
+  echo "$0: making MFCC features for low-resolution speed-perturbed data"
+  steps/make_mfcc.sh --cmd "$train_cmd" --nj $nj data/${train_set}_sp || exit 1;
+  steps/compute_cmvn_stats.sh data/${train_set}_sp || exit 1;
+  utils/fix_data_dir.sh data/${train_set}_sp
+fi
+
+if [ $stage -le 2 ]; then
+  echo "$0: aligning with the perturbed low-resolution data"
+  steps/align_fmllr.sh --nj $nj --cmd "$train_cmd" \
+    data/${train_set}_sp data/lang $gmm_dir $ali_dir || exit 1
+fi
+
+if [ $stage -le 3 ]; then
+  # Create high-resolution MFCC features (with 40 cepstra instead of 13).
+  # this shows how you can split across multiple file-systems.
+  echo "$0: creating high-resolution MFCC features"
+  mfccdir=data/${train_set}_sp_hires/data
+  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $mfccdir/storage ]; then
+    utils/create_split_dir.pl /export/fs0{1,2}/$USER/kaldi-data/mfcc/mini_librispeech-$(date +'%m_%d_%H_%M')/s5/$mfccdir/storage $mfccdir/storage
+  fi
+
+  for datadir in ${train_set}_sp ${test_sets}; do
+    utils/copy_data_dir.sh data/$datadir data/${datadir}_hires
+  done
+
+  # do volume-perturbation on the training data prior to extracting hires
+  # features; this helps make trained nnets more invariant to test data volume.
+  utils/data/perturb_data_dir_volume.sh data/${train_set}_sp_hires || exit 1;
+
+  for datadir in ${train_set}_sp ${test_sets}; do
+    steps/make_mfcc.sh --nj $nj --mfcc-config conf/mfcc_hires.conf \
+      --cmd "$train_cmd" data/${datadir}_hires || exit 1;
+    steps/compute_cmvn_stats.sh data/${datadir}_hires || exit 1;
+    utils/fix_data_dir.sh data/${datadir}_hires || exit 1;
+  done
+fi
+
+if [ $stage -le 4 ]; then
+  echo "$0: computing a subset of data to train the diagonal UBM."
+  # We'll use about a quarter of the data.
+  mkdir -p exp/nnet3${nnet3_affix}/diag_ubm
+  temp_data_root=exp/nnet3${nnet3_affix}/diag_ubm
+
+  num_utts_total=$(wc -l <data/${train_set}_sp_hires/utt2spk)
+  num_utts=$[$num_utts_total/4]
+  utils/data/subset_data_dir.sh data/${train_set}_sp_hires \
+     $num_utts ${temp_data_root}/${train_set}_sp_hires_subset
+
+  echo "$0: computing a PCA transform from the hires data."
+  steps/online/nnet2/get_pca_transform.sh --cmd "$train_cmd" \
+      --splice-opts "--left-context=3 --right-context=3" \
+      --max-utts 10000 --subsample 2 \
+       ${temp_data_root}/${train_set}_sp_hires_subset \
+       exp/nnet3${nnet3_affix}/pca_transform
+
+  echo "$0: training the diagonal UBM."
+  # Use 512 Gaussians in the UBM.
+  steps/online/nnet2/train_diag_ubm.sh --cmd "$train_cmd" --nj $nj \
+    --num-frames 700000 \
+    --num-threads 8 \
+    ${temp_data_root}/${train_set}_sp_hires_subset 512 \
+    exp/nnet3${nnet3_affix}/pca_transform exp/nnet3${nnet3_affix}/diag_ubm
+fi
+
+if [ $stage -le 5 ]; then
+  # Train the iVector extractor.  Use all of the speed-perturbed data since iVector extractors
+  # can be sensitive to the amount of data.  The script defaults to an iVector dimension of
+  # 100.
+  echo "$0: training the iVector extractor"
+  steps/online/nnet2/train_ivector_extractor.sh --cmd "$train_cmd" --nj $nj \
+     data/${train_set}_sp_hires exp/nnet3${nnet3_affix}/diag_ubm \
+     exp/nnet3${nnet3_affix}/extractor || exit 1;
+fi
+
+
+if [ $stage -le 6 ]; then
+  # We extract iVectors on the speed-perturbed training data after combining
+  # short segments, which will be what we train the system on.  With
+  # --utts-per-spk-max 2, the script pairs the utterances into twos, and treats
+  # each of these pairs as one speaker; this gives more diversity in iVectors..
+  # Note that these are extracted 'online'.
+
+  # note, we don't encode the 'max2' in the name of the ivectordir even though
+  # that's the data we extract the ivectors from, as it's still going to be
+  # valid for the non-'max2' data, the utterance list is the same.
+
+  ivectordir=exp/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires
+  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $ivectordir/storage ]; then
+    utils/create_split_dir.pl /export/fs0{1,2}/$USER/kaldi-data/ivectors/mini_librispeech-$(date +'%m_%d_%H_%M')/s5/$ivectordir/storage $ivectordir/storage
+  fi
+
+
+  # having a larger number of speakers is helpful for generalization, and to
+  # handle per-utterance decoding well (iVector starts at zero).
+  temp_data_root=${ivectordir}
+  utils/data/modify_speaker_info.sh --utts-per-spk-max 2 --respect-speaker-info $respect_speaker_info \
+    data/${train_set}_sp_hires ${temp_data_root}/${train_set}_sp_hires_max2
+
+  steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $nj \
+    ${temp_data_root}/${train_set}_sp_hires_max2 \
+    exp/nnet3${nnet3_affix}/extractor $ivectordir
+
+  # Also extract iVectors for the test data, but in this case we don't need the speed
+  # perturbation (sp).
+  for data in $test_sets; do
+    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $nj \
+      data/${data}_hires exp/nnet3${nnet3_affix}/extractor \
+      exp/nnet3${nnet3_affix}/ivectors_${data}_hires
+  done
+fi
+
+exit 0