Z9RTM/FASTDIAR-GGUF

FASTDIAR GGUF

This is a format conversion of the large checkpoint of FASTDIAR for a fork of speech.cpp, Z9RTM/speech.cpp. It is not an official release of Microsoft or of speech.cpp.

Source	microsoft/fastdiar at f48bff53b9a6a0cce4539e94b6def95f609eaea4, release v0.1.0
Checkpoint SHA-256	f6d71444eb513654a4b8c6563d11196d19283759884d4986488af3a57361228d
Weights	F32, unchanged apart from layout
Also carried	Silero VAD 16 kHz (silero_vad.jit SHA-256 e1122837f4154c511485fe0b9c64455f7b929c96fbb8d79fbdb336383ebd3720)

Conversion expands grouped downsampling weights with exact zeros, leaves out BatchNorm's training counters and adds Silero VAD's weights, so that one file holds what the diarizer runs. reference/fastdiar/convert.py in the fork does it.

Use

speech diarize FASTDIAR-large-10.4M-F32.gguf --format rttm audio.wav
speech asr haas-ja-v1.0 --vad silero-vad --diarize FASTDIAR-large-10.4M-F32.gguf --live --format json

Requires the fork's speech.cpp 0.10.2 or later.

Checks

On Apple M1 Pro with Metal, labels matched the official streaming implementation in every frame on VoxConverse clfcg and cwbvu. Windows/Vulkan is unchecked.

Limits

• One speaker per frame, so overlapped speech counts as an error.
• The number of speakers is not limited. The paper evaluates 2 to 21 speakers per recording.
• Japanese was not in the paper's evaluation and has not been measured here.
• Memory grows with the length of the recording, since the clustering keeps its whole history.

License

MIT. See LICENSE for FASTDIAR and NOTICE for ReDimNet2 (Palabra.ai) and Silero VAD.