README.md

---
license: mit
datasets:
- mozilla-foundation/common_voice_16_0
language:
- be
---

# GlowTTS + HifiGAN Male Belarusian Voice #1

This is my third attempt at training a Belarusian voice using [Coqui TTS](https://docs.coqui.ai/en/dev/index.html) and Mozilla's [CommonVoice](https://commonvoice.mozilla.org/en) dataset.  This model was developed based on the [excellent recipe](https://github.com/coqui-ai/TTS/tree/dev/recipes/bel-alex73) provided by bel-alex73. For this particular model, I tweaked the search results to find single speakers with over 30 hours of audio and selected speakers based on clarity and relatively slow speaking cadence.  This was a manual selection process that involved me tweaking bel-alex73  `choose_speaker.ipynb` notebook to show/process more that just the top ranked speaker.

This model is generated from the following client_id: 235555b6d6c6b4d882a5a0e6160f245c03e61d266c112dc3cecaeb7bcf9802d70be375ffaf9590dd7b24e95284ce06ee295da529cebd9c67f29db31cb8f092cb

I am not a native speaker of Belarusian and I am doing this to assist in my language learning efforts.  I am open to any and all feedback (esp. from native speakers) so feel free to post questions/comments.

## Sythesizing text to speech

Input text needs to be phoneme-ized in order for this model to process the speech correctly.  This process has been documented in [bel-alex73's README](https://github.com/coqui-ai/TTS/tree/dev/recipes/bel-alex73#prepare-to-training---locally).

```
tts --text "<phonemes>" --out_path output.wav \
	--config_path config.json \
	--model_path best_model.pth \
	--vocoder_config_path vocoder_config.json \
	--vocoder_path vocoder_best_model.pth
```