README.md

Myanmar Text-to-Speech Demo

This Hugging Face Space demonstrates the Myanmar Text-to-Speech (TTS) system developed by hpbyte. It's an end-to-end speech synthesis system specifically designed for the Burmese language.

⚠️ IMPORTANT: Model Files Required ⚠️

To use this demo, you must upload the following model files to the trained_model directory:

1. checkpoint_latest.pth.tar - The trained model checkpoint
2. hparams.yml - The hyperparameters configuration file

These files can be obtained from the original repository or by training the model yourself.

Configuration Instructions

Setting up the model files:

1. Create the directory structure:

   mkdir -p trained_model
   

2. Download and place model files:

   • Download checkpoint_latest.pth.tar and hparams.yml from the release page
   • Place both files in the trained_model directory

3. Verify directory structure:

   myanmar-tts-demo/
   ├── app.py
   ├── requirements.txt
   ├── trained_model/
   │   ├── checkpoint_latest.pth.tar
   │   └── hparams.yml
   └── README.md
   

4. Environment setup:

   • Make sure the PyTorch version matches the one used for training (check requirements.txt)
   • The system requires NVIDIA GPU support for optimal performance

Testing your configuration:

After setting up the model files, run a quick test to verify everything is working correctly:

from synthesis import synthesize_text

# Test with a simple phrase
test_text = "မင်္ဂလာပါ"
audio_output = synthesize_text(test_text)

# If no errors occur, your configuration is correct

About the Project

This is an implementation of Tacotron 2 for Myanmar language text-to-speech synthesis. Unlike Meta's Massively Multilingual Speech (MMS) Burmese TTS, this project is specifically focused on high-quality Burmese speech synthesis using an end-to-end approach.

Key Features:

• End-to-end Burmese text-to-speech synthesis
• Built on the Tacotron 2 architecture
• Custom text processing for the Myanmar language
• Clean and natural-sounding voice output

How to Use This Demo

1. First-time setup:

   • Upload the model files as mentioned above to the trained_model directory
   • Wait for the Space to rebuild

2. Using the demo:

   • Enter Burmese text in the input box
   • Click "Submit" to generate speech
   • Listen to the generated audio output

Examples

Try these example phrases:

• မင်္ဂလာပါ (Hello)
• မြန်မာစကားပြောစနစ်ကို ကြိုဆိုပါတယ် (Welcome to the Myanmar speech system)
• ဒီစနစ်ဟာ မြန်မာစာကို အသံအဖြစ် ပြောင်းပေးနိုင်ပါတယ် (This system can convert Myanmar text to speech)

Troubleshooting

If you encounter issues:

1. Model files not found: Make sure you've uploaded the required model files to the trained_model directory.
2. App errors: Try restarting the Space after uploading the model files.
3. Installation issues: Check the Space logs for specific error messages.
4. GPU not detected: This model requires GPU acceleration. Ensure your Hugging Face Space has GPU access enabled.
5. Module import errors: Check that all dependencies are properly installed via the requirements.txt file.

Technical Details

This model uses the Tacotron 2 architecture to generate mel spectrograms from text, which are then converted to waveforms using a vocoder. The model implementation is based on the work by hpbyte.

System Architecture:

• Text Frontend: Processes Myanmar Unicode text
• Acoustic Model: Tacotron 2 based sequence-to-sequence model with attention
• Vocoder: WaveRNN or Griffin-Lim algorithm for waveform generation

References

• Original Repository: https://github.com/hpbyte/myanmar-tts
• Paper: Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions

Credits

If you use this model in your research or application, please cite the original repository:

@misc{myanmar-tts,
  author = {Htet Pyie Sone},
  title = {Myanmar Text-to-Speech},
  year = {2021},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/hpbyte/myanmar-tts}}
}