cstr/openvoice2-tcc-GGUF

OpenVoice V2 tone-color converter — F32-sensitive GGUF

A converted checkpoint for CrispASR v0.8.48 or later. This is a tone-color converter, not a standalone text-to-speech engine. No new target voice, reference recording or validation audio is included.

openvoice2-tcc-f32sens.gguf is 102,123,872 bytes; SHA-256: 483fd233035bb2c7cde69519d4d460742b6a70c96c5293e878fe25e72d944255. The reference encoder, posterior encoder, flows and one-dimensional tensors retain F32; other decoder tensors use F16. Weight normalization is fused. The bundled base-speaker embeddings are the official upstream source tensors; no new target embedding is added. Full input hashes, converter source and validation links are in PROVENANCE.json.

Source and license

The original MyShell OpenVoice V2 model declares MIT. The unchanged MyShell MIT notice is included, from the pinned original source. This conversion changes storage and precision; it does not train a new voice. Melo and BERT are separate models: their licenses remain independent, including BERT's Apache-2.0 notice. This repository does not bundle them or the original Python installation's speech/transcription helper dependencies.

The model license does not establish rights in a particular person's identity or in an operator's reference recording. Supply an authorized reference and retain CrispASR's required generated-audio markings. Test preset identity stays unknown; the validation audio and reference captures remain private.

Validated use and limits

Use the converter with the qualified Melo V2/F16 BERT bundle. The native CLI finds this exact converter filename beside the Melo model; the Session/C ABI accepts its path through the codec setter. Both routes require an authorized reference voice and use the actual source-speaker key. The converter outputs 22.05 kHz PCM; the unconverted Melo baseline is 44.1 kHz.

Original-reference stage/magnitude checks pass on CPU and T4 CUDA. The complete current application passes both reference cases, exact reseeded repeats, observed adjacent F16 BERT discovery and all four unchanged Session/CLI speech and speaker cases. See the pinned CPU/CUDA receipts linked in PROVENANCE.json. Reference and flow work run on CPU; converter vocoder CUDA work is observed.

Qualification is bounded to the documented synthetic English cases. Broader languages, natural recordings, other GPU families and automatic VAD selection need separate validation. These results do not grant voice identity rights.