OpenMOSS/MOSS-Transcribe-Diarize

What it solves

MOSS-Transcribe-Diarize is designed to convert long-form, multi-speaker audio recordings—such as meetings, podcasts, and interviews—into structured transcripts. It eliminates the need to stitch together separate Automatic Speech Recognition (ASR) and speaker diarization systems by performing both tasks jointly in a single pass.

How it works

The model uses an end-to-end architecture combining a Whisper-Medium audio encoder and a Qwen3-0.6B style causal decoder. It processes raw audio in 30-second chunks, using a temporal merge and MLP adaptor to bridge audio features into the text backbone. The model then generates a compact transcript where each segment is marked with start/end timestamps and speaker labels (e.g., [S01], [S02]).

Who it’s for

It is intended for developers and researchers working with audio content, specifically those needing reliable transcription and speaker identification for long, messy recordings across 50+ languages.

Highlights

• End-to-End Processing: Performs transcription, diarization, and timestamping in one pass.
• High Performance: Outperforms several proprietary models (like GPT-4o and Gemini) on various objective benchmarks for Character Error Rate (CER).
• Flexible Serving: Supports deployment via SGLang Omni, vLLM, and integration with the FunASR ecosystem.
• ** uma Subtitle Tooling**: Includes a web app for uploading audio/video, reviewing segments, and exporting subtitles in JSON, SRT, or ASS formats.
• Customizable: Supports custom prompts and hotwords to improve transcription accuracy for specific terms.