Badnyal/grt-tts-base

Garo TTS – Bootstrap VITS

This is a bootstrap multi-speaker Garo TTS model trained on noisy public speech data. It is intended as an acoustic prior, not a production-quality TTS.

Notes

• Background noise may be present
• Pronunciation may be inaccurate
• Suitable for fine-tuning with clean, single-speaker studio data

Model

• Architecture: VITS
• Sample rate: 16 kHz
• Speakers: 6
• Phonemizer: eSpeak (English proxy)

Files

• best_model.pth
• config.json
• speakers.json

Inference (Python)

from TTS.utils.synthesizer import Synthesizer

synth = Synthesizer(
    tts_checkpoint="best_model.pth",
    tts_config_path="config.json",
    voice_dir=".",
    use_cuda=False,
)

wav = synth.tts(
    text="ia nokni rongde rimit rong onga",
    speaker_name="southgarohills_male"
)

synth.save_wav(wav, "output.wav")