We align speech corpora with corresponding text transcripts from scratch, focusing on letter-wise alignment rather than the more common word-wise alignment. In our approach, we replace the traditional Grapheme-to-Phoneme (G2P) process with a Grapheme-to-Grapheme (G2G) process.
We release the following components -
- Pronunciation Dictionary: Custom dictionary mapping each grapheme to itself. Contains all the letters of the language
- Acoustic Model: A model trained using the Speech Corpus and G2G Pronunciation Dictionary that extracts features and predicts phonemes.
Additionally, you will need a speech corpus: which contains audio files and corresponding text transcripts[processed].
Find the releases of all 22 Indian Languages at : https://github.com/AI4Bharat/IndicMFA/releases
The Training Details for all the languages in the release can be found in the table below:
| Language | Hours of data | IndicTTS | IndicVoices-R | Limmits | GoogleCrowdSourced | Rasa |
|---|---|---|---|---|---|---|
| Assamese | 302.86 | Y | Y | Y | ||
| Bengali | 313.32 | Y | Y | Y | Y | |
| Bodo | 222.53 | Y | Y | Y | ||
| Dogri | 91.52 | Y | Y | |||
| Gujarati | 42.9 | Y | Y | |||
| Hindi | 255.33 | Y | Y | Y | ||
| Kannada | 193.81 | Y | Y | Y | Y | Y |
| Kashmiri | 90.53 | Y | ||||
| Konkani | 85.69 | Y | ||||
| Maithili | 200.81 | Y | ||||
| Malayalam | 197.08 | Y | Y | Y | ||
| Manipuri | 55.19 | Y | Y | |||
| Marathi | 212.58 | Y | Y | Y | Y | |
| Nepali | 217.04 | Y | Y | |||
| Odia | 131.72 | Y | Y | |||
| Punjabi | 123.62 | Y | Y | |||
| Sanskrit | 76.36 | Y | ||||
| Santali | 121.84 | Y | ||||
| Tamil | 300.48 | Y | Y | Y | ||
| Telugu | 261.77 | Y | Y | Y | Y | |
| Urdu | 101.79 | Y | ||||
| Sindhi | 10.31 | Y |
To install and use the MFA software. Follow the steps at : https://montreal-forced-aligner.readthedocs.io/en/latest/index.html
In Order to get details on how to install MFA and more, go to Montreal-Forced-Aligner-Docs/README.md
For more detailed instruction on implementation, refer to training.md and inference.md.
- Python: For preprocessing scripts.
- Montreal Forced Aligner (MFA): For validation, training, and alignment.
- Praat: For analyzing TextGrid files.
- MFA Documentation: https://github.com/MontrealCorpusTools/mfa-models
- Praat Software: https://www.fon.hum.uva.nl/praat/