myG2P

Myanmar (Burmese) Language Grapheme to Phoneme Converter for automatic speech recognition (ASR) and text-to-speech (TTS) using myG2P Dictionary from the repository by ye-kyaw-thu.

Lincense

Creative Commons Attribution-NonCommercial-Share Alike 4.0 International (CC BY-NC-SA 4.0) License
Details Info of License

Introduction

Training process in deep learning approaches to ASR and TTS has become more end-to-end over a past few decades (eg. Tacotron2), eliminating the need to extract intricate linguistic features from the text. However, since this simiplicity instead requires a large amount of data, it can be a bottleneck for low-resource languages like Burmese. Recent TTS approaches that are claimed to yield human-level quality like NaturalSpeech also still require phonemes as one of their inputs instead of transcripts. Indeed, the G2P converter inside ver2 folder is developed to be used as an upstream module for NaturalSpeech, but both G2P converters provided here can also be used for any other relevant purposes.

Directory structure

Here are a few important files:

myG2P/
├── crfsuite/
│ ├── chunking.py
├── scripts/
│ ├── preprocess.sh
│ ├── evaluate.sh
│ ├── predict.py
├── tutorial/
│ ├── tutorial.pdf
├── ver1/
│ ├── g2p.model
│ ├── myg2p.ver1.txt
├── ver2/
│ ├── g2p.model
│ ├── myg2p.ver2.0.txt
├── README.md

• crfsuite: G2P converters here are trained using a popular graphical model for sequence lebeling named CRF (Conditional Random Field), and a CRF toolkit named CRFSuite is used to apply the model.
• scripts: Running preprocess.sh will create training and testing data ready to be immediately used for training the model. Running evaluate.sh will output the accuracy of the model or its prediction for the testing data depending on the option provided. Running predict.py will allow users to input any Burmese text and output its phoneme.
• ver1: g2p.model is a functioning model trained on myg2p.ver1.txt that outputs phonemes given a sentence or words. The phonemes are spelt in English intuitively, which is not a standardized way and somewhat similar to Myanglish (Burmese Language but spelled out using the English alphabet).

  သဘာဝပတ်​ဝန်းကျင်​၏ ဆန့်ကျင်ဘက်အရာသည် လုပ်ယူဖန်တီးထားသော ပတ်​ဝန်းကျင်​ဖြစ်​သည်​။ -> dha- ba wa. bjo_ wun_ kyin gyin za- kyin be' a- ja dhi lou' ju hpan ti_ hta_ dho_ ba- wun_ dha- din_ ga_
  

• ver2: g2p.model is a functioning model trained on myg2p.ver2.0.txt that outputs phonemes given a sentence or words. The phonemes are spelt in IPA (International Phonetic Alphabet), which is the ideal format for NaturalSpeech.

  သဘာဝပတ်​ဝန်းကျင်​၏ ဆန့်ကျင်ဘက်အရာသည် လုပ်ယူဖန်တီးထားသော ပတ်​ဝန်းကျင်​ဖြစ်​သည်​။ -> ðə bà wa̰ pʰó wʊ́ɴ zə gá zə tɕɪ̀ɴ bɛʔ ʔə jà ðì loʊʔ jù pʰàɴ tí tʰá ðɔ́ bə wʊ́ɴ kə lə baʔ
  

Installation and Usage

Procedures for installing CRFSuite, training and evaluation are documented in this notion page. Some procedures from the original tutorial are automated and some installation steps for CRFSuite have been modified to reflect the current state of the software distribution.

Acknowledgement

The development of these Burmese G2P converters can be largely attributable to the G2P dictionary for training and tutorials provided by Mr. Ye Kyaw Thu, and the repository hosting them can be found here.

Future Developments

• Incorporating sentence level conversion (Current model predicts incorrect phonemes for characters not included in nouns, verbs, and adjectives such as particles and post-positional markers)
• Changing burmese digits into words