Problem in building clustergen indic voice #13
Comments
|
Hi, This looks like an issue with labeling or utts creation. Do you have a log of the previous steps so that I can better figure this out? Also the steps you pointed out are a bit old. I can point you at the latest set of steps. |
|
Here is the series of steps to build a decent voice assuming Festival, Festvox, speechtools and SPTK are installed (let me know if this is an issue): Setup the directory structure
(16 July 2020 so so so sorry I entered this wrong. The command should be as follows: setup_cg_indic NOT setup_cg sorry for the inconvenience) Copy the wavefiles and prompts./bin/get_wavs ${LOCATION}/*.wav Some Text Processing./bin/do_build build_prompts etc/txt.done.data Feature Extraction./bin/do_clustergen parallel f0_v_sptk etc/txt.done.data Combining the features for Machine Learningmv festvox/clustergen.scm festvox/clustergen.scm.xxx Separate train and test splits./bin/traintest etc/txt.done.data Training./bin/do_clustergen parallel cluster etc/txt.done.data.train Testing./bin/do_clustergen cg_test resynth cgp etc/txt.done.data.test |
|
Thanks for the such detailed help. I followed the latest set of steps suggested. Got stuck with the following issues on the similar sample of 100 wav files from 'cmu_indic_hin_ab.tar.bz2'.: ######################## Issue 1############################ ######################## Issue 2############################ ######################## Issue 3############################ hindi_0003 COMBINE_COEFFS (f0,mcep_deltas,str,v) ######################## Issue 4############################ Not gone further after this much of error...... |
|
Seems to be an issue with versions.
What is the version of festival you are using?
And sptk
…On 19 Aug 2018 5:32 a.m., "shrikant6153" ***@***.***> wrote:
Thanks for the such detailed help.
I followed the latest set of steps suggested. Got stuck with the following
issues on the similar sample of 100 wav files from
'cmu_indic_hin_ab.tar.bz2'.:
######################## Issue 1############################
./bin/do_build build_prompts etc/txt.done.data
SIOD ERROR: could not open file ./festvox/language_variant.scm
closing a file left open: ./festvox/indic_lexicon.scm
closing a file left open: ./festvox/cmu_indic_hin_lexicon.scm
closing a file left open: festvox/cmu_indic_hin_clunits.scm
closing a file left open: festvox/build_clunits.scm
(Note:- This was there before too in the earlier scripts. But I had
resolved it by putting a file named "language_variant.scm" with content
'hin' in 'cmu_indic_hin_ab/festvox/' directory.
######################## Issue 2############################
./bin/do_clustergen parallel str_stpk etc/txt.done.data # Strengths of
excitation
do_clustergen: unknown options str_stpk tmpdir/dobuild_parallelworker.
31140.4
do_clustergen: unknown options str_stpk tmpdir/dobuild_parallelworker.
31139.3
do_clustergen: unknown options str_stpk tmpdir/dobuild_parallelworker.
31137.1
do_clustergen: unknown options str_stpk tmpdir/dobuild_parallelworker.
31141.5
do_clustergen: unknown options str_stpk tmpdir/dobuild_parallelworker.
31138.2
do_clustergen: unknown options str_stpk tmpdir/dobuild_parallelworker.
31136.0
do_clustergen: unknown options str_stpk tmpdir/dobuild_parallelworker.
31142.6
do_clustergen: unknown options str_stpk tmpdir/dobuild_parallelworker.
31155.7
do_clustergen: unknown options str_stpk tmpdir/dobuild_parallelworker.
31173.10
do_clustergen: unknown options str_stpk tmpdir/dobuild_parallelworker.
31171.9
do_clustergen: unknown options str_stpk tmpdir/dobuild_parallelworker.
31177.11
do_clustergen: unknown options str_stpk tmpdir/dobuild_parallelworker.
31186.13
do_clustergen: unknown options str_stpk tmpdir/dobuild_parallelworker.
31168.8
do_clustergen: unknown options str_stpk tmpdir/dobuild_parallelworker.
31183.12
do_clustergen: unknown options str_stpk tmpdir/dobuild_parallelworker.
31187.14
do_clustergen: unknown options str_stpk tmpdir/dobuild_parallelworker.
31204.15
######################## Issue 3############################
./bin/traintest etc/txt.done.data
hindi_0003 COMBINE_COEFFS (f0,mcep_deltas,str,v)
hindi_0002 COMBINE_COEFFS (f0,mcep_deltas,str,v)
hindi_0004 COMBINE_COEFFS (f0,mcep_deltas,str,v)
hindi_0016 COMBINE_COEFFS (f0,mcep_deltas,str,v)
hindi_0001 COMBINE_COEFFS (f0,mcep_deltas,str,v)
hindi_0005 COMBINE_COEFFS (f0,mcep_deltas,str,v)
hindi_0006 COMBINE_COEFFS (f0,mcep_deltas,str,v)
hindi_0007 COMBINE_COEFFS (f0,mcep_deltas,str,v)
cat: str/hindi_0001.str: No such file or directory
cat: str/hindi_0002.str: No such file or directory
cat: str/hindi_0004.str: No such file or directory
cat: str/hindi_0006.str: No such file or directory
cat: str/hindi_0005.str: No such file or directory
cat: str/hindi_0007.str: No such file or directory
cat: str/hindi_0016.str: No such file or directory
cat: str/hindi_0003.str: No such file or directory
..............
issue3.txt <https://github.com/festvox/festival/files/2300204/issue3.txt>
######################## Issue 4############################
./bin/do_clustergen cg_test resynth cgp etc/txt.done.data.test
Error reading ESPS file /home/shrikant/festival_hindi_
tts/indic/cmu_indic_hin_ab//festival/trees/cmu_indic_hin_mcep.params
Cannot load track: /home/shrikant/festival_hindi_
tts/indic/cmu_indic_hin_ab//festival/trees/cmu_indic_hin_mcep.params
SIOD ERROR: could not open file /home/shrikant/festival_hindi_
tts/indic/cmu_indic_hin_ab//festival/trees/cmu_indic_hin_mcep.tree
awk: cmd. line:2: fatal: division by zero attempted
awk: cmd. line:2: fatal: division by zero attempted
awk: cmd. line:2: fatal: division by zero attempted
awk: cmd. line:2: fatal: division by zero attempted
Not gone further after this much of error......
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#13 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AV3IMUuOTK5KHhreQ45-lW3QUojHE9Kjks5uSTDBgaJpZM4WCooo>
.
|
|
same as mentioned in the script fest_build. I just ran the script and then source the "export_various_PATHS.sh" in order to export them. |
|
This seems the issue. I see that you have latest versions. Just to be sure, can you create a new directory and run only the prompt building command ( ./bin/do_build build_prompts) that gave this error. Let me know if this happens again |
|
shall I put the content of file ./festvox/language_variant.scm as hin? |
|
Issues 2 through 4 are caused by 1. The content of /festvox/language_variant.scm should be 'hin' by default. |
|
Note that the sample build script in fest_build.txt is for English. When building an indic voice, the command to setup directory structure is: |
|
Hello saikrishna! I followed exactly as instructed. Please see, I attached complete directory in a zip. |
|
One more point! During the execution of step13, a continuous stream of multiple segmentation fault errors are coming: For other steps (1 --> 12), out files are there in the above attached zipped folder. Not able to figure out where the problem is actually because I dont see any difference in errors earlier and now. Thanks for the support and help saikrishna btw. |
|
For some reason I am unable to download the directory. Can you attach ou1, out2 here |
|
Nothing wrong with these. I next am suspecting some issue in feature extraction. Can you attach out6 , out7, out8. |
|
The problem is with out8. There are two things we can do for this: (1) Ignore this and continue voice building. In this case, modify the next step to the following from combine_coeffs_me uses strengths of excitation In this case, we also need to modify the clustergen.scm file and indicate that we are not using mixed excitation. An easy way to do this is the following: cp festvox/clustergen.scm.xxx festvox/clustergen.scm # ( We previously made an explicit in this file through steps 9 and 10 that we would be using Mixed excitation. So we are just reverting.) Now you can run the clustering step: (2) The other (and real) solution is to dig deeper into why str_sptk is failing. Can you paste the file ./bin/do_clustergen here so that I can inspect it. It should support the argument 'str_sptk' |
|
For (2), Bear if you can the do_clustergen file is here (1) Let me incorporate it. |
|
Wait. I just noticed that the spelling is incorrect in the step8 in the script you shared. It should be str_sptk not str_stpk |
|
Once things run smoothly till 'cluster', I'd say run the duration model, the following without 'parallel' instead of ./bin/do_clustergen parallel dur etc/txt.done.data.train |
|
I realize that I made that spelling error when I shared the steps. Sorry for that :) |
|
Still the same situation. Please have a look again what I did this time.
Got the ouput files: |
|
Step 11 has an error on the last phone z_3. Can you run that step again. Once that runs successfully, should be fine. I also notice that there are seg faults in step 13 log. Segmentation fault might also be occuring due to less space being allocated. There is a parameter called SIODHEAPSIZE in do_clustergen. Increasing that should alleviate this fault. |
|
Step 11 you mean to say: this command I checked : SIODHEAPSIZE=20000000 in the ./bin/do_clustergen file |
|
@step 11 yes Just double the heap size and see ( not multiplying by 10). This is usually not necessary tbh |
|
With updated SIODHEAPSIZE=25000000 |
|
Hi, I was able to download the zip. When I ran the step, it did finish without any issues. Here are the last lines from log: RMSE 0.1516 Correlation is 0.9867 Mean (abs) Error 0.0963 (0.1171) I was able to finish the duration model ( next step) and generate test samples too. This is weird since I am essentially continuing from your folder structure |
|
@awbcmu Can you look into this |
|
Given the failure of the missing language_variant.scm file I suspect initialization with the wrong version might be the culprit. Also note it should be str_sptk not str_stpk as the option. Another suggestion it running with the parallel option. If you run out of memory and something dies, that might be hard to detect in the next step. I would regenerate the templates, and then copy in the waveforms and txt.done.data |
|
@shrikant6153 can you run it without 'parallel' once |
|
with or without parallel : -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- |
|
cool. keep me posted :) also try building a voice ignoring the corpus building process. the only thing is perhaps try to avoid super long sentences. Try building a voice with around 2 to 3 hours and check if ownas has a decent count. closing for now. reopen here / new issue and tag me if something goes wrong |
|
@saikrishnarallabandi I just trained around 12 hours of data. with strength of excitation from your comment I must say while corpus creation I allowed 40000 words while this process to make lexicon, The default was 5000 words. While training i got many of these warnings Is this because of my lexicon? I will update how well it does without pruning silences. |
|
This is the warning from duration model training. Duration modeling is the
final step in the voice building process.
This warning indicates that there is a phone / there are phones which have
less number of examples. Duration model is built at the HMM state level
within FestVox by dividing each phone into N states(mostly 3).
…On Tue, Jul 21, 2020 at 5:50 PM Prajwal Rao ***@***.***> wrote:
@saikrishnarallabandi <https://github.com/saikrishnarallabandi> I just
trained around 12 hours of data. with strength of excitation from your
comment
<#13 (comment)>
Here is the log file
whole_process_with_mixed_feature.txt
<https://github.com/festvox/festival/files/4956404/whole_process_with_mixed_feature.txt>
I must say while corpus creation I allowed 40000 words while this process
to make lexicon, The default was 5000 words.
$FESTVOXDIR/src/promptselect/make_nice_prompts make_freq_lex
While training i got many of these warnings
*festival/dur/data/dur.data.train.train: bad float -inf in field
lisp_zscore_dur vector *
Is this because of my lexicon?
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#13 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AFO4QMNOXYTG3DS6W5X6VTDR4YETBANCNFSM4FQKRIUA>
.
|
|
I have checked the logfile.
Initial labeling ran for 23 iterations. Typically it finishes before 20 for
most voices but 23 seems ok.
Feature Extraction seems ok.
Clustering seems ok.
The only issue seems to be with the duration modeling.
You should be able to generate test samples
The file test/cgp/scores should have mcd and other metrics
test/tts shud have the test samples
On Tue, Jul 21, 2020 at 5:57 PM Sai Krishna Rallabandi <
[email protected]> wrote:
… This is the warning from duration model training. Duration modeling is the
final step in the voice building process.
This warning indicates that there is a phone / there are phones which have
less number of examples. Duration model is built at the HMM state level
within FestVox by dividing each phone into N states(mostly 3).
On Tue, Jul 21, 2020 at 5:50 PM Prajwal Rao ***@***.***>
wrote:
> @saikrishnarallabandi <https://github.com/saikrishnarallabandi> I just
> trained around 12 hours of data. with strength of excitation from your
> comment
> <#13 (comment)>
> Here is the log file
> whole_process_with_mixed_feature.txt
> <https://github.com/festvox/festival/files/4956404/whole_process_with_mixed_feature.txt>
>
> I must say while corpus creation I allowed 40000 words while this process
> to make lexicon, The default was 5000 words.
> $FESTVOXDIR/src/promptselect/make_nice_prompts make_freq_lex
>
> While training i got many of these warnings
> *festival/dur/data/dur.data.train.train: bad float -inf in field
> lisp_zscore_dur vector *
>
> Is this because of my lexicon?
>
> —
> You are receiving this because you were mentioned.
> Reply to this email directly, view it on GitHub
> <#13 (comment)>,
> or unsubscribe
> <https://github.com/notifications/unsubscribe-auth/AFO4QMNOXYTG3DS6W5X6VTDR4YETBANCNFSM4FQKRIUA>
> .
>
|
|
I just for the results of cg_test and the mean still seems to be high Will update if I'm able to synthesize any audio. |
|
Unfortunately, the synthesized data is not understandable. Attaching sample |
|
Ya i was afraid of that. the MCD is too high. If you have access to GPU based compute, here is the run script for building a Tacotron voice. The example is from Hindi speaker(male). The script creates exp/taco_one_phseq. Checkpoints and intermediate files(attention plots, etc) will be in exp/taco_one_phseq/checkpoints. The logfile will be in exp/taco_one_phseq/tracking/logfile. TB eventfiles will be in exp/taco_one_phseq/tracking. If you choose to use this, pay attention to line number 35. I am selecting the shortest 600 utterances in the example script. It might be worth playing with this number. With this configuration, you should be able to see clear attention around 10K timesteps. The model trains for ~200K time steps if you leave it running. |
|
Listening to the samples, clearly the duration model seems wrong. What about the wavefiles in the directroy test/cgp? These use original durations |
|
This is my original sample data. The data in test/cgp sounds similar to test/tts, probably a little better. Here is an example. Yes i have access to GPUs. I am open to trying Falcon, my only requirement is CPU inferencing and fast inferencing. Will get back after trying this script. |
|
The cases where i saw festvox voice not have a satisfactory output are:
1) less number of examples for some phones
2) conversational speech that has filled pauses(umms n aahhs etc)
3) wierd silences in the data. If this is the case, running the scripts
bin/prune_silence n bin/prune_middle_silences helps (try this once)
4) un natural zeros in the signal. Sometimes the recording studio adds
"zeros" to the signal. This can be mitigated by adding small random noise
to the signal.
5) if the data was recorded using a laptop n power cord is on, it adds a
50(0r 60) hz humm to the signal. This can break voice building. But am not
sure this is the case with your data.
Apart from that the process should work typically. Only exceptions i found
in standard voices were for (a) speaker aup. That voice became better with
Randon forests and (b) speaker slp(marathi) that voice became better with
falcon.
If the dataset has super long sentences the process runs for long but does
result in a decent voice.
…On Wed, 22 Jul 2020, 03:09 Prajwal Rao, ***@***.***> wrote:
This is my original sample data.
data_orig.zip
<https://github.com/festvox/festival/files/4958375/data_orig.zip>
The data in test/cgp sounds similar to test/tts, probably a little better.
Here is an example.
data_001.zip
<https://github.com/festvox/festival/files/4958349/data_001.zip>
Yes i have access to GPUs. I am open to trying Falcon, my only requirement
is CPU inferencing and fast inferencing. Will get back after trying this
script.
Im also curious to know where exactly Im going wrong.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#13 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AFO4QMLQ3L56Q5BUUMFQXEDR42GEDANCNFSM4FQKRIUA>
.
|
|
I already had pruned the silences for this dataset. Both end and middle using the scripts provided. Im not sure about points 4 and 5. Will have to check. I think the data is recorded through an app on the phone. Could there be something wrong with alignment? How can I verify this? |
|
I was slightly worried when it took 23 iterations of HMM to get
alignments(from your logfile)
the alignments can be checked by files in test/cgp.
The alignment information is saved in lab files in the lab/ directory.
Looking at lab/*.lab might help judge alignments
…On Wed, Jul 22, 2020 at 12:54 PM Prajwal Rao ***@***.***> wrote:
I already had pruned the silences for this dataset. Both end and middle
using the scripts provided.
There could actually be less number of examples of some phones. What is
the average word count required? If i pick top 5000 words my highest word
count is 7513 and lowest word count is 5.
data.wc.txt
<https://github.com/festvox/festival/files/4961314/data.wc.txt>
Im not sure about points 4 and 5. Will have to check. I think the data is
recorded through an app on the phone.
Could there be something wrong with alignment? How can I verify this?
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#13 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AFO4QMIWOQEXHLDGLUI2TC3R44KU3ANCNFSM4FQKRIUA>
.
|
|
In my case following this to build data did not work. Finally got these results for 42h of data: all mean 11.247145 std 999.729682 var 999459.436663 n 55196550
F0 mean 45.817497 std 32.325401 var 1044.931537 n 2207862
noF0 mean 0.160005 std 0.108984 var 0.011878 n 52988688
MCD mean 4.549429 std 1.591600 var 2.533191 n 2207862 |
|
Cool awesome.
Thanks for pointing it out. Will check
…On Sun, Jul 26, 2020 at 5:13 PM Prajwal Rao ***@***.***> wrote:
In my case following this <http://festvox.org/bsv/c2176.html> to build
data did not work.
But when generated the txt.data.done on my own and copied the wavs using
./bin/get_wavs, it worked!
However, both the txt.done.data was the same!
Thank you for your detailed help @saikrishnarallabandi
<https://github.com/saikrishnarallabandi>
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#13 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AFO4QMIWFSVFZNJEMKROTLDR5SL6JANCNFSM4FQKRIUA>
.
|
|
Voice compilation script/process is acting weird. I was able to compile PUNJABI voice and also associated flite voice but it failed to compile the same voice again and also failed to compile hindi voice. Most of the console errors are similar to what are mentioned in the issue from 2018. I tried to follow both the processes i.e. one mentioned in that issue comments long with do_indic command but none of it succeeds. all my festival/speech tools etc are clean build by fest_build.sh command. Intriguing part is that it compiled the voice twice without any problem but now it fails. |
|
can u send me logfile?
…On Thu, May 13, 2021 at 9:41 AM plehal ***@***.***> wrote:
Voice compilation script/process is acting weird. I was able to compile
PUNJABI voice and also associated flite voice but it failed to compile the
same voice again and also failed to compile hindi voice. Most of the
console errors are similar to what are mentioned in the issue from 2018.
I tried to follow both the processes i.e. one mentioned in that issue
comments long with do_indic command but none of it succeeds. all my
festival/speech tools etc are clean build by fest_build.sh command.
Intriguing part is that it compiled the voice twice without any problem but
now it fails.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#13 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AFO4QMORLXX5KT6WXSNEI7DTNPJJPANCNFSM4FQKRIUA>
.
|
|
What is the location of the logfile? Most of these errors are on the console where do_indic command is ran. Please, let me know if the log is captured in some file. Otherwise, I'll redirect stdout to afile and rerun the command. I ran one session this morning which is still running after more than 10 hours and seems to be going better than before just looking at correlation numbers.... Dataset of 16 vectors of 67 parameters from: wagon_rf_572210/data
|
|
Ah it doesn't automatically get saved in a file. The second option you
said, redirecting to a file manually is the way. Can you do that?
I don't need the logfile for random forest voice(which runs for longer). I
just need to see it for base build(which should get sooner than 3 hours
unless the sentences are too long). Just one thing- do this in a new
directory. I am attaching the script for base build if it helps
…On Fri, May 14, 2021, 00:36 plehal ***@***.***> wrote:
What is the location of the logfile? Most of these errors are on the
console where do_indic command is ran. Please, let me know if the log is
captured in some file. Otherwise, I'll redirect stdout to afile and rerun
the command. I ran one session this morning which is still running after
more than 10 hours and seems to be going better than before just looking at
correlation numbers....
Dataset of 16 vectors of 67 parameters from: wagon_rf_572210/data
RMSE 0.0767 Correlation is 0.9465 Mean (abs) Error 0.0516 (0.0568)
Iteration 16 festival/trees/nX_1_mcep.tree
Iteration 17 festival/trees/nX_2_mcep.tree
Dataset of 17 vectors of 67 parameters from: wagon_rf_572233/data
Attempt to access channel 53 of 52 channel track
Attempt to access channel 54 of 52 channel track
Attempt to access channel 55 of 52 channel track
Dataset of 17 vectors of 67 parameters from: wagon_rf_572233/data
RMSE 0.6720 Correlation is 0.8022 Mean (abs) Error 0.1739 (0.6497)
Dataset of 16 vectors of 67 parameters from: wagon_rf_572210/data
Attempt to access channel 52 of 52 channel track
Attempt to access channel 53 of 52 channel track
Dataset of 16 vectors of 67 parameters from: wagon_rf_572210/data
RMSE 0.5969 Correlation is 0.9498 Mean (abs) Error 0.1331 (0.5824)
Dataset of 4122 vectors of 67 parameters from: wagon_rf_580592/data
Attempt to access channel 52 of 52 channel track
Attempt to access channel 53 of 52 channel track
Attempt to access channel 55 of 52 channel track
Dataset of 4122 vectors of 67 parameters from: wagon_rf_580592/data
RMSE 0.5827 Correlation is 0.9535 Mean (abs) Error 0.1320 (0.5676)
RMSE 0.1190 Correlation is 0.9910 Mean (abs) Error 0.0739 (0.0933)
Iteration 11 festival/trees/i:_3_mcep.tree
Iteration 13 festival/trees/hv_3_mcep.tree
Dataset of 8487 vectors of 67 parameters from: wagon_rf_580258/data
Attempt to access channel 52 of 52 channel track
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#13 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AFO4QMMN5UGYZUEWBHJH7P3TNSSEDANCNFSM4FQKRIUA>
.
|
|
Thanks. Sure, I'll rerun it and send you the log file. Really appreciate the responsiveness. |
|
Here is the log file. (gzipped). It took more than 24 hours to complete the run which ended in a failure.(indic pan amp) |
|
Hi,
The log file shows Random forest build process. Just to confirm,\
(1) Did you run this in a new directory(which looks likely)
(2) Did you use the script basebuild_indic I attached? (this doesnt seem
likely)
The basebuild generates a much smaller log file and runs faster too(~1 hour)
From the current log file, its harder to debug since it includes random
forest build, but as far as I can see, it looks ok till feature extraction.
The alignment seems alright. Can you confirm you used the script i
attached? or if not, use the script to generate log file?
I am attaching it again in this. It should finish in an hour or so
…On Sun, May 16, 2021 at 8:33 PM plehal ***@***.***> wrote:
Here is the log file. (gzipped). It took more than 24 hours to complete
the run which ended in a failure.(indic pan amp)
p.log.gz <https://github.com/festvox/festival/files/6490358/p.log.gz>
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#13 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AFO4QMMNC4OOYYMUJB4RZI3TOBP4BANCNFSM4FQKRIUA>
.
|
|
No, this is the resulting log file for "do_indic pan amp" script. OK, let me run your base build script and catch the log. |
|
The script does create a festvox/festival voice. Now, how do I create flite voice from it as this script does create the basic directory structure for flite. Thanks for the help. |
|
Awesome
Try these lines
# Build a flite voice from this build
rm -rf flite
$FLITEDIR/tools/setup_flite
./bin/build_flite cg
cd flite
make
…On Wed, May 19, 2021, 01:15 plehal ***@***.***> wrote:
The script does create a festvox/festival voice. Now, how do I create
flite voice from it as this script does create the basic directory
structure for flite. Thanks for the help.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#13 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AFO4QMPZ4PP27FBNZ7B2SP3TONCO3ANCNFSM4FQKRIUA>
.
|
|
Thanks. The flite file is built. However, it doesn't work, whereas the corresponding festvox voice does work. I'll open a new issue here to keep this thread sane. |
Following the link "http://festvox.org/bsv/x3528.html", tried to built hindi tts from scratch on sample of100 'hindi' wav files ['hindi_0001.wav' - 'hindi_0102.wav'] obtained from 'cmu_indic_hin_ab.tar.bz2'.
Every script works fine upto the following script--
./bin/do_clustergen parallel cluster etc/txt.done.data.train
getting lots of 'file not found' & 'Segmentation fault (core dumped)' Errors
example------------ final lines of above command:
Segmentation fault (core dumped)
Dataset of 0 vectors of 64 parameters from: festival/feats/9r_1.feats
Attempt to access channel 1 of 0 channel track
Dataset of 0 vectors of 64 parameters from: festival/feats/9r_1.feats
Segmentation fault (core dumped)
Segmentation fault (core dumped)
Dataset of 0 vectors of 64 parameters from: festival/feats/9r=_1.feats
Attempt to access channel 1 of 0 channel track
Dataset of 0 vectors of 64 parameters from: festival/feats/9r=_1.feats
Dataset of 0 vectors of 64 parameters from: festival/feats/9r_1.feats
Attempt to access channel 1 of 0 channel track
Dataset of 0 vectors of 64 parameters from: festival/feats/9r_1.feats
Segmentation fault (core dumped)
Dataset of 0 vectors of 64 parameters from: festival/feats/9r=_1.feats
Attempt to access channel 1 of 0 channel track
Dataset of 0 vectors of 64 parameters from: festival/feats/9r=_1.feats
Segmentation fault (core dumped)
Dataset of 0 vectors of 64 parameters from: festival/feats/9r_1.feats
Attempt to access channel 1 of 0 channel track
Dataset of 0 vectors of 64 parameters from: festival/feats/9r_1.feats
Segmentation fault (core dumped)
Segmentation fault (core dumped)
Dataset of 0 vectors of 64 parameters from: festival/feats/9r=_1.feats
Attempt to access channel 1 of 0 channel track
Dataset of 0 vectors of 64 parameters from: festival/feats/9r=_1.feats
Dataset of 0 vectors of 64 parameters from: festival/feats/9r_1.feats
Attempt to access channel 1 of 0 channel track
Dataset of 0 vectors of 64 parameters from: festival/feats/9r_1.feats
Segmentation fault (core dumped)
Dataset of 0 vectors of 64 parameters from: festival/feats/9r=_1.feats
Attempt to access channel 1 of 0 channel track
Dataset of 0 vectors of 64 parameters from: festival/feats/9r=_1.feats
Segmentation fault (core dumped)
Dataset of 0 vectors of 64 parameters from: festival/feats/9r_1.feats
Attempt to access channel 1 of 0 channel track
Dataset of 0 vectors of 64 parameters from: festival/feats/9r_1.feats
Segmentation fault (core dumped)
Dataset of 0 vectors of 64 parameters from: festival/feats/9r=_1.feats
Attempt to access channel 1 of 0 channel track
Dataset of 0 vectors of 64 parameters from: festival/feats/9r=_1.feats
Segmentation fault (core dumped)
Segmentation fault (core dumped)
Dataset of 0 vectors of 64 parameters from: festival/feats/9r_1.feats
Attempt to access channel 1 of 0 channel track
Dataset of 0 vectors of 64 parameters from: festival/feats/9r_1.feats
Segmentation fault (core dumped)
Segmentation fault (core dumped)
Segmentation fault (core dumped)
Segmentation fault (core dumped)
Segmentation fault (core dumped)
Segmentation fault (core dumped)
Segmentation fault (core dumped)
Collect trees
SIOD ERROR: wrong type of argument to setcar
BACKTRACE:
0: (set-car! (car tree) vector_num)
1: (clustergen::dump_tree_vectors tree rawtrackfd)
2: (set! tree (clustergen::dump_tree_vectors tree rawtrackfd))
3: (f (car l2))
4: (cons (f (car l2)) r)
5: (set! r (cons (f (car l2)) r))
6: (while l2 (set! r (cons (f (car l2)) r)) (set! l2 (cdr l2)))
7: (mapcar
(lambda
(unit)
(...)
...)
unittypes)
8: (if
(consp cg:multimodel)
(mapcar
(...)
cg:multimodel)
...)
9: (begin
(set! cg:parallel_tree_build t)
(build_clustergen "etc/txt.done.data.train"))
closing a file left open: festival/trees/cmu_indic_ss_mcep.rawparams
closing a file left open: festival/trees/cmu_indic_ss_mcep.tree
Please tell the solution where I am doing wrong.
Note:- I have build all the required tools as mentioned in 'fest_build' script.
The text was updated successfully, but these errors were encountered: