FST based Malayalam Phonetic Analyser

A detailed note by Kavya Manohar.

What is a Phonetic analyser?

‘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.

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 helpful to transliterate the  word to a different script. It will be useful if the phonetic  representation can be converted back to the grapheme sequence -  requirement for speech synthesis systems. A finite state transducer (FST) helps us to achieve this.

Finite State Transducers consists of a finite number of states which are linked by transitions  labelled 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.

Grapheme to Phoneme (g2p) mapping

FSTs can be used for mapping graphemes to phonemes and the reverse.  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 schwa 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.

The phonetic representation currently used is the International  Phonetic Alphabets (IPA). Along with IPA, the articulatory features of  the consonants are provided. I have used Thunchath Ezhuthachan Malayalam University’s phonetic archive as a reference for the mapping. The mapping can be further extended to phonetic alphabets suitable for various TTS systems.

SFST and HFST Tool-kit

My g2p implementation is based on Stuttgart Finite State Transducer (SFST) and Helsinki Finite-State Technology (HFST).  Both are programming language for finite state transducers which is  based on extended regular expressions with variables. Written in SFST  programming paradigm, the code is compiled to create an automata. HFST’s  wider programming interfaces utilise this automata to provide python  api, web api and command line interface for Malayalam phonetic analyser.

Web Demo

You can now easily test the Malayalam phonetic analyser system here:

It analyses single Malayalam words and displays IPA and phonetic details. Source code and conversations can be found on project page.

Future Works

Currently,a transliteration of graphemes to IPA along with  articulatory details of consonants as ‘tags’ is done by this analyser.  This won’t be sufficient for a speech synthesis system. Spoken Malayalam  has a lot of contextual variations in phonemes from what is currently  implemented. A thorough study on the same and its implementation is an  immediate future plan.

Most TTS systems use their own internal phonetic representation.  Mapping of IPA to such systems can be done, if it can be directly  converted to speech by such systems.

References

1. Open morphology for Finnish
2. Malyalam morphological analyser using finite state transducers
3. The Festvox Indic Frontend for Grapheme-to-Phoneme Conversion
4. Malayalam Phonetic Archive by Thunchath Ezhuthachan Malayalam University
5. IPA and sounds

This article was originally written by Kavya Manohar and published at kavyamanohar.com