You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
v0.8.50 — Mini-BS-RoFormer, SCNet, MeanVC2, and correctness fixes
Source separation
Mini-BS-RoFormer V2.5 (HiDolen) native separation: four stems, F16/Q8_0/Q4_K, auto-download via -m mini-bs-roformer (cstr/mini-bs-roformer-GGUF). CLI --separate, HTTP server and Session (#562).
On Apple Metal it currently runs on the CPU, because the Metal path loses most of the signal (#568). Set CRISPASR_MINI_BS_ROFORMER_METAL=1 to opt back in for debugging.
SCNet and SCNet-large native separation (#566): strict 372-tensor conversion, F16/Q8_0/Q4_K, CLI/server/Session. Qualified against the original Python model on every stage, waveform and speed gate for both sizes. The upstream MIT notice and the author's permission for the checkpoints are retained.
Recognition
MOSS-Transcribe-Diarize: Session and CLI now forward the beam size and offer optional seeded sampling (temperature → top-k → top-p, in HF order); greedy stays the default (#564). New CLI aliases --top-k and --top-p.
moss-audio: encodes only the real audio instead of padding every clip to 30 s, so time markers and event times on short clips are correct (#569). CRISPASR_MOSS_AUDIO_PAD30=1 restores the old path.
Haas-ja-v1.0 with the KDA predictor in the parakeet runtime (#561).
parakeet-unified-en-0.6b: the converter now takes the model's own tokenizer (#567). The published GGUFs were already corrected in place; re-download if your copy decodes to nonsense.
Language identification
LID models are no longer downloaded silently. A missing LID model is fetched only with --auto-download or after a TTY prompt; otherwise LID is skipped with a hint (#570). moss-audio with --prompt skips LID, since the result would be unused.
Voice conversion and synthesis
MeanVC2 streaming zero-shot voice conversion, with C ABI, registry and bindings.
ArkTTS Metal acceleration and codebook clamping.
Performance
CPU speedups across mel/fbank/LFR, conformer attention, CTC alignment and beam search, wav2vec2, SenseVoice, Paraformer, FireRed, F5-TTS, VoxCPM2, Mel-Band RoFormer, MeanVC2, MioCodec, diarization, CQT and RVQ/iSTFT. Each comes with a paired A/B speed and quality bench and a serial fallback switch.
CosyVoice3 vocoder FASTCONV path.
Fixes
Lint Deep (cppcheck) is green again. Fixed: an ArkTTS virtual call in the destructor, a Mini-BS lambda, a parakeet null check, and the sampler comparator.
