A ggml/C++ port whose output is byte-identical to the genuine PyTorch pipeline — transcript text, every timestamp and every speaker tag. One pass over the whole file, greedy decode, f16 weights.
No windowing · no audio-KV eviction · no repetition penalty · no EOS-coverage suppression · no loop guard · no cross-window speaker linking · no s2tw conversion · no number ITN. Every one of these shipped in the previous build of this Space. They were removed because each changes the output, and the claim here is exact agreement with the reference. Long-audio windowing returns in stage 2.
These two clips are the ones the port is gated on. Their reference transcripts (produced by the genuine PyTorch model in f32) ship with this Space, so the page can diff this engine's live output against the reference in your browser and tell you whether it still matches.
Pick an example.
| reference | vs this engine (f16) |
|---|---|
| PyTorch f32 | byte-identical, CPU and CUDA, both clips |
| PyTorch bf16 | differs — bf16's 8-bit mantissa is coarser than f16's 10 |
| PyTorch f16 | differs — different accumulation order |
Engine: RapidSpeech.cpp @ moss-pure, which vendors moss-transcribe.cpp (MIT) unmodified, loaded through its C API so the model is resident and audio is passed as in-memory PCM. Weights: moss-transcribe-f16.gguf — base model, no fine-tuning. Model: OpenMOSS MOSS-Transcribe-Diarize (Apache-2.0).