MOSS-Transcribe-Diarize transcribes and attributes speech in a single pass.

MOSI open-sources MOSS-Transcribe-Diarize, a 0.9B parameter model that transcribes, aligns, and attributes multi-voice audio in a single end-to-end pass.

Signed MOSI (OpenMOSS Team), MOSS-Transcribe-Diarize is now open source: an audio comprehension model that transforms a multi-voice recording into a structured transcription, complete with speakers and timestamps. Only one version is currently open, a 0.9 billion parameter model under an Apache 2.0 license, with a more powerful Pro version expected via API.

The interest lies in its end-to-end approach. Where classic pipelines chain together three distinct components—speech recognition (ASR), diarization (assigning each segment to a speaker), and temporal alignment—similar to WhisperX, this model treats all three as a single generation task. It answers "what was said," "by whom," and "when" in one go, producing a timestamped text where each turn of speech carries a speaker label ([S01], [S02]). Under the hood, a Whisper-Medium audio encoder feeds a Qwen3-0.6B type decoder.

Compactness is the core argument. With 0.9 billion parameters and a 128,000-token context window, the model processes up to ninety minutes of audio in a single pass, without segmentation or reassembly. On a single RTX 4090 card, MOSI measures approximately 100 tokens per second, meaning a five-to-ten-minute file is transcribed in less than thirty seconds. A hotwords system accepts proper nouns, business terms, or product references in advance for better recognition.

In terms of results, the publisher claims first place on the AISHELL-4 benchmark, with an error reduction of over 40% compared to common models and speaker attribution presented as stable over long conversations. The repository includes weights, code, a technical report, and a small web subtitling application (SRT, ASS, or JSON export, embedding via FFmpeg), with day-one support from SGLang Omni, vLLM, and MLX-audio.