MOSS-Transcribe-Diarize 0.9B is an open-source, state-of-the-art, end-to-end audio understanding model. It addresses the challenge of processing long-form, multi-speaker audio by jointly performing transcription, diarization, and timestamp generation in a single pass. Designed for applications like meetings, podcasts, and interviews, it provides structured, speaker-aware transcripts with optional acoustic event annotations, offering a unified and reliable solution for audio analysis.

How It Works

The model employs an end-to-end architecture, integrating a Qwen3-0.6B style causal decoder with a Whisper-Medium encoder. Audio features are processed and fused with text embeddings, enabling the model to directly output compact, time-aligned transcripts. This unified approach avoids stitching separate ASR and diarization systems, aiming for improved reliability and a richer understanding of audio content by producing precise timestamps and consistent speaker labels (e.g., [S01]).

Quick Start & Requirements

• Installation: Clone the repository (git clone https://github.com/OpenMOSS/MOSS-Transcribe-Diarize.git), navigate to the directory, set up a Python 3.12 environment using uv venv --python 3.12 .venv, activate it, and install dependencies with uv pip install -e ".[torch-runtime]" --torch-backend=auto.
• Prerequisites: Python 3.12, Transformers 5.x, PyTorch (implied by torch-runtime).
• Links: Official GitHub repository (implied by clone URL), prompt examples (examples/prompts.md).

Highlighted Details

• Achieves state-of-the-art (SOTA) performance on various audio understanding benchmarks.
• Provides end-to-end transcription, diarization, timestamps, and acoustic event awareness.
• Supports efficient serving via vLLM (OpenAI-compatible API) and SGLang Omni.
• Includes a local web application for subtitle generation, review, and export (SRT, ASS, JSON), with optional FFmpeg burn-in.

Maintenance & Community

The project is developed by the OpenMOSS team. No specific community channels (e.g., Discord, Slack) or details on core contributors or sponsorships are provided in the README.

Licensing & Compatibility

The project is described as "open-source." Specific license details and compatibility notes for commercial use or closed-source linking are not explicitly stated in the provided README.

Limitations & Caveats

A more powerful "Pro" version is announced for future API access. For processing very long audio files, the max_new_tokens parameter may require significant increases (e.g., to 65536) to ensure complete transcription. Performance benchmarks are reported on single-H100 hardware.