README.md

metadata

license: mpl-2.0
language:
  - en
  - pt
library_name: phoonnx
pipeline_tag: text-to-speech
tags:
  - tts
  - yourtts
  - vits
  - onnx
  - phoonnx
  - offline
  - voice-cloning

phoonnx-yourtts

Mirror of a YourTTS checkpoint exported to ONNX for phoonnx, OpenVoiceOS's offline TTS library. YourTTS is a multilingual VITS conditioned on an external 512-d speaker d-vector rather than a speaker-id embedding, which is what enables zero-shot voice cloning: a reference clip is run through a bundled speaker encoder (coqui_resnet) to produce the d-vector that conditions synthesis. See phoonnx/engines/yourtts.py and docs/cloning.md.

This is the original multilingual YourTTS model (Casanova et al., 2022), trained by Coqui on VCTK (English) and a Brazilian Portuguese corpus, with speaker d-vectors for a fixed set of training speakers exposed as per-speaker config files.

Note: unlike the other phoonnx-* mirrors, this repository is not currently wired into phoonnx's packaged voice_index — there is no your-tts/* entry shipped by update-cache/list-voices yet. The files are present and the yourtts engine adapter is implemented and tested; load them by path as shown below until a voice-index entry is published.

Contents

• your-tts/model.onnx — the shared YourTTS ONNX graph (all speakers below use this one model)
• configs/your-tts-en-spk0.json, -spk1.json, -spk2.json — three bundled English (VCTK) training speakers, each with its own d-vector in engine_params
• configs/your-tts-pt-spk0.json, -spk1.json, -spk2.json — three bundled Brazilian Portuguese training speakers

Usage

import wave
from huggingface_hub import hf_hub_download
from phoonnx.voice import TTSVoice

model_path = hf_hub_download("OpenVoiceOS/phoonnx-yourtts", "your-tts/model.onnx")
config_path = hf_hub_download("OpenVoiceOS/phoonnx-yourtts", "configs/your-tts-en-spk0.json")

voice = TTSVoice.load(model_path, config_path)
with wave.open("out.wav", "wb") as wav_file:
    voice.synthesize_wav("Hello, this is a test.", wav_file)

Zero-shot cloning from your own reference clip (overrides the bundled speaker) works the same way, passing speaker_reference to synthesize_wav — see docs/cloning.md.

License

YourTTS's pretrained weights were released by Coqui as part of the coqui-ai/TTS model zoo under the MPL-2.0 license, the same license as the Coqui TTS toolkit itself (this predates Coqui's later CPML non-commercial license, introduced with XTTS). phoonnx itself is Apache-2.0.