Meari accent base — GPT-SoVITS v2ProPlus sound stage (ONNX) and voice values
Shared part of the regional accent reader in Meari, a local speech engine (MIT). Each region adds its own package with the fine-tuned text encoder and GPT stage (for example Miyauel/meari-accent-gyeongsang). Runs with ONNX Runtime on the CPU, without PyTorch.
| File | What |
|---|---|
vits.onnx |
SoVITS sound stage of GPT-SoVITS v2ProPlus (s2Gv2ProPlus.pth from lj1995/GPT-SoVITS), exported by Meari (engine/spikes/dialect/gsv_onnx_export.py). Not fine-tuned — the same for every region |
voices/<voice>.npz |
Precomputed reference values for Meari's 10 built-in voices: phones of the reference text, semantic tokens, spectrogram and speaker embedding. The reference sentences were spoken by the Supertonic 3 preset voices; the values were computed once with GPT-SoVITS's feature models, so the app does not need those models |
Licenses
- GPT-SoVITS code and pretrained weights: MIT (RVC-Boss/GPT-SoVITS). The authors do not disclose the training data of the pretrained weights.
- Voice values (
voices/): derived from the Supertonic 3 preset voices, whose weights are under the OpenRAIL-M license — its use restrictions apply (for example: no impersonation, disclose that speech is AI-generated). - The semantic tokens were extracted with
chinese-hubert-base(MIT; trained on WenetSpeech, whose terms limit use to non-commercial purposes). The HuBERT model itself is not included.