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 and Mozilla's CommonVoice dataset. This model was developed based on the excellent recipe 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.
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