README.md

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

# GlowTTS + HifiGAN Female Belarusian Voice #2

This is my second 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: 24addf14d781fed7a0b038ce1cfa5c1341bea14eccdef21272fdb78b017d28e8464882fc010f9f19c77ba5d55ea06a6451ec0accd16a72255da2e74168387ae0

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
```