‘Phoneme’ is the fundamental unit in the the speech system of the language. ‘Grapheme’ is the fundamental unit in the writing system. From one or more graphemes a phoneme can be synthesized. A phonetic analyser analyses the written form of the text to give the phonetic characteristics of the grapheme sequence.
Using Virtual Environment (https://docs.python.org/3/library/venv.html) is recommended.
To start using this python library
pip install mlphon
The following python snippet will split a word in Malayalam script into syllables.
from mlphon import PhoneticAnalyser
mlphon = PhoneticAnalyser()
mlphon.split_to_syllables('കേരളം')
It will give the result
['കേ', 'ര', 'ളം']
from mlphon import PhoneticAnalyser
mlphon = PhoneticAnalyser()
mlphon.analyse('കേരളം')
It gives the result as a sequence of ipa and associated phonetic tags.
[[{'phonemes': [{'ipa': 'k',
'tags': ['plosive', 'voiceless', 'unaspirated', 'velar']},
{'ipa': 'eː', 'tags': ['v_sign']}]},
{'phonemes': [{'ipa': 'ɾ', 'tags': ['flapped', 'alveolar']},
{'ipa': 'a', 'tags': ['inherentvowel']}]},
{'phonemes': [{'ipa': 'ɭ', 'tags': ['lateral', 'retroflex']},
{'ipa': 'a', 'tags': ['inherentvowel']},
{'ipa': 'm', 'tags': ['anuswara']}]}]]
from mlphon import PhoneticAnalyser
mlphon = PhoneticAnalyser()
mlphon.grapheme_to_phoneme('കാറ്റ്')
It gives the ipa sequence as output.
['kaːṯṯə']
from mlphon import PhoneticAnalyser
mlphon = PhoneticAnalyser()
mlphon.phoneme_to_grapheme('kaːṯṯə')
It gives the corresponding grapheme sequences as output.
['കാറ്റ്']
usage: mlphon [-h] [-s] [-a] [-p] [-pe string] [-se string] [-g] [-i INFILE]
[-o OUTFILE] [-v]
optional arguments:
-h, --help show this help message and exit
-s, --syllablize Syllablize the input Malayalam string
-a, --analyse Phonetically analyse the input Malayalam string
-p, --tophoneme Transcribe the input Malayalam grapheme to phoneme
sequence
-pe string, --phoneme_end string
String to be inserted at end of phoneme
-se string, --syllable_end string
String to be inserted at end of syllable
-g, --tographeme Transcribe the input phoneme sequence to Malayalam
grapheme
-i INFILE, --input INFILE
source of analysis data
-o OUTFILE, --output OUTFILE
target of generated strings
-v, --verbose print verbosely while processing
For example to perform g2p operation on a set of words stored in input.txt with one Malayalam word per line,
mlphon -p -pe " " -se "."-i path/to/inputfile.txt -o path/to/outputfile.txt
Inputfile contents:
cat path/to/inputfile.txt
അകത്തുള്ളത്
അകപ്പെട്ടത്
അകലെ
Outputfile contents:
അകത്തുള്ളത് a .k a .t̪ t̪ u .ɭ ɭ a .t̪ ə .
അകപ്പെട്ടത് a .k a .p p e .ʈ ʈ a .t̪ ə .
അകലെ a .k a .l e .
A typical use case of phonetic analysis is to create a phonetic lexicon to be used in Automatic Speech Recognition or Text to Speech Synthesis. The phonetic representation with each phoneme separated by a space can be obtained as below:
from mlphon import PhoneticAnalyser, split_as_phonemes
mlphon = PhoneticAnalyser()
split_as_phonemes(mlphon.analyse('ഇന്ത്യയുടെ')[0])
It results in the output:
'i n̪ t̪ j a j u ʈ e'
The phonetic representation with each syllable separated by a space can be obtained as below:
from mlphon import PhoneticAnalyser, split_as_syllables
mlphon = PhoneticAnalyser()
split_as_syllables(mlphon.analyse('ഇന്ത്യയുടെ')[0])
It results in the output:
'i n̪t̪ja ju ʈe'
To get phonemes and syllables with user defined end-marker strings as below:
from mlphon import PhoneticAnalyser, phonemize
mlphon = PhoneticAnalyser()
analysis = mlphon.analyse('ഇന്ത്യയുടെ')
for result in analysis: # To loop through multiple analysis results, if any
phonemize(result, " ", ".")
It results in the output with a 'space' after every phoneme and a 'period' after every syllable:
'i .n̪ t̪ j a .j u .ʈ e .'
Understanding the phonetic characteristics of a word is helpful in many computational linguistic problems. For instance, translating a word into its phonetic representation is needed in the synthesis of a text to speech (TTS) system. The phonetic representation is needed to transliterate the word to a different script. It will be more useful if the phonetic representation can be converted back to the grapheme sequence. A finite state transducer (FST) helps us to achieve this.
Finite State Transducers provide a method for performing mathematical operations on ordered collections of context-sensitive rewrite rules such as those commonly used to implement fundamental natural language processing tasks. Multiple rules may be composed into a single pass, mega rule, significantly increasing the efficiency of rule-based systems. An FST consists of a finite number of states which are linked by transitions labeled with an input/output pair. The FST starts out in a designated start state and jumps to different states depending on the input, while producing output according to its transition table.
In this project we try to develop a phonetic analyser for malayalam script. A specific application of transliterating malayalam script to international phonetic alphabet (IPA) is demonstrated. Specifically, the system is developed using Stuttgart Finite State Toolkit(SFST) formalism.
FSTs when applied to GPC systems, the mapping is between the graphemes of the writing system of of a language and phonemes of the speech of that language. This transducer can be implemented as the composition of different transducers, each performing a specific mapping task. The whole task can be implemented by FST chains- One FST for rule based grapheme-phoneme mapping, another FST for implementing inherentvowel addition depending on the context and so on.
Grapheme to phoneme (g2p) correspondence may not be always one-to-one. If the orthography (writing system) of a language is phonemic, then its g2p conversion would have been straightforward. Malayalam, like other indic languages has mostly phonemic orthography unlike English which is non-phonemic. The g2p mapping of malayalam requires certain contextual rules to be applied to handle inherentvowel addition at beginning/end/middle of words depending on the presence of chillus and virama, phonetic changes that occur in the context of certain sequence of consonants, contextual nasalisation etc. It is usually required that the process is bidirectional. Ie., the grapheme to phoneme correspondence (GPC) system should be able to retrieve the orthographic representation of the language in the native script from the phonetic sequence. Malayalam GPC using FST
The chain of transducers used din this system and their function are listed below:
$wordfilter$
This is the first level transducer which accepts malayalam scripts and add and word wrapping tags before further processing.
TODO:Provisions to accept punctuation marks, malayalam/arabic numerals, archaic malayalam characters, latin text etc. Currently it assumes what is given as input is a word. It will act as word splitter in future
$syllable$
It splits syllables with tags to indicate beginning and end of syllables respectively for words input to it with and tags
$syllablizer$ is a composition of $wordfilter$ and $syllable$
g2p mapping is done on syllable splitted words. So $syllablizer$ is a prerequisite for g2p processing. $g2p$ is composed of the followings FSTs
$IPAmap$
This transducer accepts inputs from the output of previous transducer and performs the IPA mapping. During this process along with associating graphemes to phonemes, tags are added to indicate if it is a pure vowel, a vowel sign or a consonant. The tags added by this transducer are:
<virama> <vowel> <v_sign> <visaraga> <anuswara> <velar> <palatal> <retroflex> <dental> <alveolar> <labial> <labiodental> <glottal> <chil> <plosive> <voiceless> <unaspirated> <voiced> <aspirated> <nasal> <fricative> <flapped> <lateral> <approximant> <glide> <trill>
The malayalam script assumes every consonant if not followed by a virama, has the inherent vowel associated with it. But this FST does not associate the inherent vowel to every consonant. But presence of a virama is clearly indicated using a tag <virama> for further processing. Only atomic chillus are accepted by the system and <chil> tag added.
$inherentvowel$
Inherent vowel has to be added to all consonants if it is at <EoS> or when it is followed by <anuswara> or <visarga>.
This context is identified and inherentvowel addition is done along with a <inherentvowel> tag.
TODO: Presence of any special character including space, period, comma, exclamation mark etc to be identified and inherentvowel addition to be done. Inherent vowel takes a special for certain graphemes at the <BoW>. This has to be handled.Eg- രമ്യ - രെമ്യ , ഇല - എല
$tta_nta$
The unicode sequence റ+ ് + റ has a special phonetic mapping (ṯṯ) which is different from the phonetic representation (r) of റ .
Similar is the case with ന + ് + റ . Its phonetic mapping is (nṯ) which is much different from the mapping of ന(n̪) or റ(r).
This stage of FST replaces the already mapped റ+ ് + റ r<trill><alveolar><virama>r<trill><alveolar> to ṯ<plosive><voiceless><unaspirated><alveolar><virama>ṯ<plosive><voiceless><unaspirated><alveolar> and ന + ് + റ n̪<nasal><dental><virama>r<trill><alveolar> to n<nasal><alveolar><virama>ṯ<plosive><voiceless><unaspirated><alveolar>.
$reph$
The reph symbol in Malayalam corresponding to ് + ര (്ര) follows other consonants. But such conjuncts like ക്ര, ത്ര, സ്ര, ശ്ര etc. have the pronunciation which is closer to റ (alveolar trill) rather than ര (flapped trill). reph.fst has contextual rules to replace <virama>ɾ<flapped> with<virama>r<trill>.
$rephexp$
But there is an exception to the above rule for the conjuncts ഗ്ര and ദ്ര.
$schwa$
Adds the half-u or Samvruthokara, whenever there is a virama at word ends.
eg: അവന് /aʋanə/ is different from അവൻ /aʋan/
$alveolarnasal$
Performs the disambiguation of ന - as alveolar or dental nasal. Contextual rules are written, for complementary pair contexts. Works well: പന, നായ, നനവ്, ആന, അന്നു, ഇന്നലെ, സന്നിധാനം, ഊനം, ഊന്നൽ, വിഭിന്നം,
There are exceptions to general phonological contexts which can be addressed only by morphological parser and POS tags. Fails: മന്നൻ, അനുനയം,
TODO: ഭംഗി -> ഭങ്ങി , ചിഹ്നം -> ചിന്നം
$analysis$ represents the overall FST which combines each of the above FSTs in a chain.
$ml2ipa$ is the fst that converts the Malayalam script to IPA in the analysis mode. It uses the $g2p$ FST combined with a tag filter $tagfilter$ to achieve this. But tags like <visarga> <zwnj> are explicitly retained in the IPA sequence.
You need Stuttgart Finite-State Transducer Technology (SFST) (https://www.cis.uni-muenchen.de/~schmid/tools/SFST/) to compile this analyzer. The Makefile provided compiles all the sources and produces the binary FSA analysis.a, ml2ipa.a, syllablizer.a.
In a debian/ubuntu based GNU/Linux, SFST can be installed as follows
$ sudo apt install sfst
Clone or download this git repository to your machine.
$ make
- Open morphology for Finnish https://github.com/flammie/omorfi
- Malyalam morphological analyser using finite state transducers https://github.com/santhoshtr/mlmorph
- The Festvox Indic Frontend for Grapheme-to-Phoneme Conversion https://www.parlikar.com/files/aup_wildre_2016.pdf
- Malayalam Phonetic Archive by THUNCHATH EZHUTHACHAN MALAYALAM UNIVERSITY http://www.cmltemu.in/phonetic/#/
- IPA and sounds http://www.internationalphoneticalphabet.org/ipa-sounds/ipa-chart-with-sounds/
If you are using Mlphon in your project, please cite this paper
@ARTICLE{9877808,
author={Manohar, Kavya and Jayan, A. R. and Rajan, Rajeev},
journal={IEEE Access},
title={Mlphon: A Multifunctional Grapheme-Phoneme Conversion Tool Using Finite State Transducers},
year={2022},
volume={10},
number={},
pages={97555-97575},
doi={10.1109/ACCESS.2022.3204403}}