MOSS-Transcribe-Diarize: transcribe.cpp GGUF

GGUF conversions of OpenMOSS-Team/MOSS-Transcribe-Diarize for use with transcribe.cpp.

Ported from upstream commit d7231bb, pinned 2026-07-12. Validated against the MOSS author repo (OpenMOSS/MOSS-Transcribe-Diarize) reference at transcribe.cpp commit 3f5e15c on 2026-07-12.

Offline English/Chinese speech-to-text with speaker diarization. A 0.9B audio-LLM: a Whisper-Medium encoder (24 layers, d_model=1024) feeds a 4x temporal merge + VQAdaptor bridge into a Qwen3-0.6B decoder (28 layers) via audio-token injection. Takes a 16 kHz mono WAV and emits transcript text in the canonical diarized format [start][Sxx]text[end], where the speaker tags and segment timestamps are generated text, not special tokens. Not a streaming model.

Downloads

WER measured on the full LibriSpeech test-clean split (2620 English utterances) with the Whisper-style English text normalizer and jiwer 3.x. MOSS emits the diarized format [start][Sxx]text[end]; the bracket spans are metadata and are de-diarized to a space (for both hypothesis and reference) before scoring, matching the author-repo reference runner. The same-manifest MOSS author-repo reference (bf16, greedy) lands at 2.07% with 95% bootstrap CI [1.82%, 2.40%]. The BF16 port lands at 2.08% (within +0.01 of the reference, well inside the CI band); the lower-bit presets sit between 1.93% and 1.99% (statistical noise) except Q4_K_M at 2.59%, whose excess is a handful of 4-bit tail failures (6 empty outputs, 5 English->Chinese language-drift utterances, 1 timestamp-token repetition loop) rather than broad degradation. Prefer Q5_K_M or higher if those tail failures matter. Reproduce with scripts/wer/run.py + scripts/wer/score.py.

Usage

Build transcribe.cpp from source:

git clone git@github.com:handy-computer/transcribe.cpp.git
cd transcribe.cpp
cmake -B build && cmake --build build

Run on a 16 kHz mono WAV:

build/bin/transcribe-cli \
  -m MOSS-Transcribe-Diarize-Q8_0.gguf \
  input.wav

If your audio isn't already 16 kHz mono WAV, convert it first:

ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav

See the transcribe.cpp model page for performance numbers, numerical validation, and reproduction steps.

License

Inherited from the base model: Apache-2.0. See the upstream model card for full terms.

---

Original Model Card

The section below is reproduced from OpenMOSS-Team/MOSS-Transcribe-Diarize at commit d7231bb for offline reference. The upstream card is the authoritative source.

MOSS-Transcribe-Diarize

MOSS-Transcribe-Diarize 0.9B is an end-to-end audio understanding model for long-form multi-speaker transcription, diarization, timestamps, and acoustic event awareness.

Given an audio or video file, the model generates a compact speaker-aware transcript in one pass, including timestamps and anonymous speaker labels such as [S01], [S02], and beyond.

News

• 2026-07-09: Released MOSS-Transcribe-Diarize 0.9B.

Contents

• Introduction
• Model Architecture
• Evaluation
• Quickstart
  • Environment Setup
  • Python Usage
  • Custom Prompt and Hotwords
  • Serve with vLLM and SGLang
  • Subtitle Web App
• Output Format
• More Information
• License
• Citation

Introduction

MOSS-Transcribe-Diarize 0.9B turns real-world long-form audio into structured, speaker-aware transcripts in one pass. Instead of stitching together separate ASR and diarization systems, it jointly performs speech transcription and speaker diarization, producing time-aligned text with consistent speaker labels.

The model is built for meetings, calls, podcasts, interviews, lectures, videos, and other long or messy multi-speaker recordings. It can also emit acoustic event annotations, giving downstream systems a richer view of what happened, who spoke, and when.

Core capabilities:

• Long-form transcription: Converts long audio or video recordings into timestamped text.
• Speaker-aware diarization: Assigns anonymous speaker labels such as [S01] and [S02] without a separate diarization pipeline.
• Promptable generation: Supports custom transcription instructions, hotwords, and acoustic event annotations.

Model Architecture

This Hugging Face repository includes the custom Transformers remote code required to load the model with trust_remote_code=True.

Evaluation

We evaluate MOSS-Transcribe-Diarize using three objective metrics: Character Error Rate (CER), concatenated minimum-permutation Character Error Rate (cpCER), and Delta-cp. Lower is better for all metrics. A dash (-) indicates that the result is unavailable.

Quickstart

Environment Setup

Use a clean Python environment. The model uses custom Transformers code, so load the model and processor with trust_remote_code=True.

conda create -n moss-transcribe-diarize python=3.12 -y
conda activate moss-transcribe-diarize

git clone https://github.com/OpenMOSS/MOSS-Transcribe-Diarize.git
cd MOSS-Transcribe-Diarize

pip install --index-url https://download.pytorch.org/whl/cu128 torch torchaudio
pip install -e .

The GitHub package provides helper utilities such as audio/video loading, transcription message construction, transcript parsing, CLI inference, and the subtitle web app. The model weights and remote-code model files are loaded from this Hugging Face repository.

Python Usage

import torch
from transformers import AutoModelForCausalLM, AutoProcessor

from moss_transcribe_diarize import parse_transcript
from moss_transcribe_diarize.inference_utils import (
    build_transcription_messages,
    generate_transcription,
    resolve_device,
)

model_id = "OpenMOSS-Team/MOSS-Transcribe-Diarize"
audio_path = "audio.wav"

device = resolve_device("auto")
dtype = torch.bfloat16 if device.type == "cuda" else torch.float32

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype="auto",
).to(dtype=dtype).to(device).eval()

processor = AutoProcessor.from_pretrained(
    model_id,
    trust_remote_code=True,
)

messages = build_transcription_messages(audio_path)
result = generate_transcription(
    model,
    processor,
    messages,
    max_new_tokens=2048,
    do_sample=False,
    device=device,
    dtype=dtype,
)

print(result["text"])

for segment in parse_transcript(result["text"]):
    print(segment.start, segment.end, segment.speaker, segment.text)

The message flow follows the common Qwen multimodal pattern:

1. processor.apply_chat_template(messages, tokenize=False) renders text with audio placeholders.
2. The helper utilities load audio waveforms from the same messages.
3. processor(text=text, audio=audios) computes Whisper input features and expands audio placeholders.
4. model.generate(...) produces timestamped transcription and diarization text.

Custom Prompt and Hotwords

The default prompt is optimized for timestamped transcription and speaker diarization:

请将音频转写为文本，每一段需以起始时间戳和说话人编号（[S01]、[S02]、[S03]…）开头，正文为对应的语音内容，并在段末标注结束时间戳，以清晰标明该段语音范围。

To add hotwords, append a short hint to the default prompt:

请将音频转写为文本，每一段需以起始时间戳和说话人编号（[S01]、[S02]、[S03]…）开头，正文为对应的语音内容，并在段末标注结束时间戳，以清晰标明该段语音范围。热词提示：热词1, 热词2, 热词3

More prompt recipes are available in the GitHub repository: https://github.com/OpenMOSS/MOSS-Transcribe-Diarize/blob/main/examples/prompts.md

Serve with vLLM and SGLang

MOSS-Transcribe-Diarize supports vLLM serving through the OpenAI-compatible transcription API. Use a pinned vLLM nightly build that includes the MOSS-Transcribe-Diarize model registration. Choose one of the following commands: for CUDA 12 environments, use cu129; for CUDA 13 environments, use cu130.

uv pip install -U vllm \
  --torch-backend=auto \
  --extra-index-url https://wheels.vllm.ai/68b4a1d582818e67adc903bf1b8fc5a5447da2fa/cu129

or:

uv pip install -U vllm \
  --torch-backend=auto \
  --extra-index-url https://wheels.vllm.ai/68b4a1d582818e67adc903bf1b8fc5a5447da2fa/cu130

vllm serve OpenMOSS-Team/MOSS-Transcribe-Diarize --trust-remote-code

curl http://localhost:8000/v1/audio/transcriptions \
  -F model="OpenMOSS-Team/MOSS-Transcribe-Diarize" \
  -F file=@"audio.wav" \
  -F response_format="json" \
  -F temperature="0"

The recommended way to serve MOSS-Transcribe-Diarize is SGLang Omni through the OpenAI-compatible /v1/audio/transcriptions endpoint. Install sglang-omni by following the installation guide, then download the model:

hf download OpenMOSS-Team/MOSS-Transcribe-Diarize

Serve the model:

sgl-omni serve \
  --model-path OpenMOSS-Team/MOSS-Transcribe-Diarize \
  --port 8000 \
  --max-running-requests 16 \
  --cuda-graph-max-bs 16 \
  --mem-fraction-static 0.80

Use response_format=verbose_json when you need parsed speaker segments. json returns the raw transcript text only.

curl -X POST http://localhost:8000/v1/audio/transcriptions \
  -F model=OpenMOSS-Team/MOSS-Transcribe-Diarize \
  -F file=@audio.wav \
  -F response_format=verbose_json

import requests

with open("audio.wav", "rb") as f:
    resp = requests.post(
        "http://localhost:8000/v1/audio/transcriptions",
        data={
            "model": "OpenMOSS-Team/MOSS-Transcribe-Diarize",
            "response_format": "verbose_json",
        },
        files={"file": ("audio.wav", f, "audio/wav")},
        timeout=300,
    )

resp.raise_for_status()
payload = resp.json()
print(payload["text"])
for segment in payload.get("segments", []):
    print(f"[{segment['start']:.2f}-{segment['end']:.2f}] {segment['text']}")

For longer multi-speaker audio, raise max_new_tokens so the decoder can finish the full diarized transcript:

curl -X POST http://localhost:8000/v1/audio/transcriptions \
  -F model=OpenMOSS-Team/MOSS-Transcribe-Diarize \
  -F file=@audio.wav \
  -F response_format=verbose_json \
  -F max_new_tokens=65536

For benchmarking, performance numbers, and implementation details, see the SGLang Omni cookbook. The following single-H100 results are reported for short- and long-sequence multi-speaker ASR tasks.

movies short-sequence ASR:

aishell4_long long-sequence ASR:

Subtitle Web App

The source package includes a local subtitle workflow for upload, review, subtitle export, and optional FFmpeg burn-in:

mtd-subtitle-web \
  --model OpenMOSS-Team/MOSS-Transcribe-Diarize \
  --host 127.0.0.1 \
  --port 7860

Open http://127.0.0.1:7860, upload an audio/video file, review the parsed subtitle segments, then download JSON/SRT/ASS or burn an MP4 if ffmpeg and ffprobe are available on PATH.

For batch processing:

mtd-subtitle /path/to/input.mp4 \
  --model OpenMOSS-Team/MOSS-Transcribe-Diarize \
  --out-dir runs/example \
  --render

Output Format

The canonical output format is:

[start_time][Sxx]transcribed speech[end_time]

Example:

[0.48][S01]Welcome everyone[1.66][12.26][S02]The new transcription pipeline is ready for evaluation[13.81][14.36][S01]Great, include the diarization results in the report[18.76]

In this format:

• start_time and end_time are timestamps in seconds.
• [S01], [S02], and similar labels are anonymous model-generated speaker labels.
• Speaker labels are relative labels within the input audio and should not be interpreted as real speaker identities.

More Information

• GitHub: https://github.com/OpenMOSS/MOSS-Transcribe-Diarize
• MOSI.AI: https://mosi.cn
• OpenMOSS: https://www.open-moss.com

License

MOSS-Transcribe-Diarize 0.9B is licensed under the Apache License 2.0.

Citation

If you use MOSS-Transcribe-Diarize 0.9B, please cite the technical report:

@misc{moss_transcribe_diarize_2026,
  title={MOSS Transcribe Diarize Technical Report},
  author={{MOSI.AI}},
  year={2026},
  eprint={2601.01554},
  archivePrefix={arXiv},
  primaryClass={cs.SD},
  url={https://arxiv.org/abs/2601.01554}
}