cstr/MOSS-Transcribe-Diarize-GGUF

MOSS-Transcribe-Diarize-0.9B GGUF

GGUF conversions of OpenMOSS-Team/MOSS-Transcribe-Diarize for CrispASR.

Joint ASR + speaker diarization + timestamps in a single 0.9B model. Produces timestamped, speaker-labelled transcripts in one pass.

Files

File	Size	Description
moss-transcribe-diarize-0.9b-f16.gguf	1.7 GB	Full precision (F16)
moss-transcribe-diarize-0.9b-q8_0.gguf	1.4 GB	8-bit quantized
moss-transcribe-diarize-0.9b-q4_k.gguf	1.2 GB	4-bit quantized (recommended)
diff-harness-ref/moss-diarize-ref.gguf	9.5 MB	Diff harness reference (jfk.wav ground truth)

Usage

crispasr --backend moss-diarize -m auto -f audio.wav -osrt

With hotwords:

crispasr --backend moss-diarize \
  -m moss-transcribe-diarize-0.9b-q4_k.gguf \
  -f meeting.wav \
  --hotwords "MOSS,CrispASR" \
  -osrt

Output Format

[00:00:00.320 --> 00:00:02.220]  (Speaker 1) And so, my fellow Americans,
[00:00:03.020 --> 00:00:07.640]  (Speaker 1) ask not what your country can do for you,
[00:00:08.110 --> 00:00:10.540]  (Speaker 1) ask what you can do for your country.
[00:00:11.440 --> 00:00:15.580]  (Speaker 2) And so, my fellow Americans, ask not...

Architecture

Stock Whisper encoder (24L, 80 mel, Conv1d) with 4x temporal merge, VQAdaptor, time markers every 5s, and Qwen3-0.6B LM decoder. Output: [timestamp][Sxx]text[timestamp] format with speaker labels.

Diff Harness

All 4 stages pass at cos=1.000 on both F32 and Q4_K:

Stage	cos_min	max_abs
mel_spectrogram	1.000000	1.20e-04
conv_stem_out	1.000000	6.11e-05
encoder_output	1.000000	1.24e-04
audio_embeds	1.000000	1.35e-02

Conversion

python models/convert-moss-transcribe-diarize-to-gguf.py \
  --input OpenMOSS-Team/MOSS-Transcribe-Diarize \
  --output moss-transcribe-diarize-0.9b-f16.gguf

crispasr-quantize moss-transcribe-diarize-0.9b-f16.gguf \
                  moss-transcribe-diarize-0.9b-q4_k.gguf q4_k

License

Apache-2.0 (same as the base model).