Piper TTS — Gujarati male
A lightweight Gujarati TTS model trained from scratch using Piper TTS, exported to ONNX (~63 MB) for fast edge deployment.
Model details
| Attribute | Value |
|---|---|
| Framework | OHF-Voice/piper1-gpl |
| Language | Gujarati (gu) |
| Speaker | Male (single speaker) |
| Format | ONNX (~63 MB) |
| Sample rate | 22050 Hz |
| Training | 144 epochs from scratch on NVIDIA L4 (24 GB) |
| Phonemizer | espeak-ng (Gujarati gu) |
Training data
Trained on AI4Bharat/Rasa Gujarati Male dataset (15,797 clips, ~26 hours, CC-BY-4.0).
This model uses only openly-licensed training data. No YouTube-scraped audio was used. Free for both commercial and non-commercial use.
Files
| File | Description |
|---|---|
epoch_144.onnx |
Final ONNX model (recommended) |
epoch_50.onnx |
Milestone — first recognizable Gujarati |
epoch_89.onnx |
Milestone — improved clarity |
epoch_139.onnx |
Milestone — near-final quality |
Usage
echo "નમસ્તે, કેમ છો?" | piper --model epoch_144.onnx --output_file output.wav
Training code
Full training pipeline and troubleshooting: BhammarArjun/TTS_3_training_edge
License
Apache 2.0 — Free for any use including commercial.
Training data (AI4Bharat Rasa) is CC-BY-4.0. No restrictively-licensed data was used.
Citation
@misc{arjun2026pipergu,
title = {Piper TTS Gujarati male — trained from scratch},
author = {Arjun Bhammar},
year = {2026},
url = {https://huggingface.co/Arjun4707/piper-gujarati-male}
}
Acknowledgements
- OHF-Voice / Piper for the TTS framework
- AI4Bharat for the Rasa Gujarati Male dataset
- espeak-ng for Gujarati phoneme support