The Year in AI Papers
The year/Independent research

Paper 2601.01554

MOSS Transcribe Diarize Technical Report

MOSI. AI, :, Donghua Yu, Zhengyuan Lin, Hanfu Chen, Chen Yang, Yiyang Zhang, Jingqi Chen, Ke Chen, Liwei Fan, Yi Jiang, Jie Zhu, et al.

Published
Jan 2026
Research lab
Independent
Citations
5
GitHub
Not linked

01 In brief

Summary

MOSS Transcribe Diarize is a unified multimodal large language model for Speaker-Attributed, Time-Stamped Transcription (SATS), jointly performing word recognition, speaker attribution, and timestamp prediction in a single end-to-end pass.

It uses a 128k-token context window to process up to 90 minutes of audio without chunking, preserving long-range speaker memory and discourse continuity.

Trained on real in-the-wild conversational audio and property-aware simulated mixtures, it outperforms state-of-the-art commercial systems (Doubao, ElevenLabs, GPT-4o, Gemini 2.5 Pro, Gemini 3 Pro) on AISHELL-4, Podcast, and Movies benchmarks, achieving the lowest CER, cpCER, and Δcp.

The model's architecture couples an audio encoder with a projection module into a pretrained text LLM, using formatted timestamp text between audio chunks for accurate temporal encoding.

Evaluation metrics include CER (ASR accuracy), cpCER (joint ASR and diarization), and Δcp (attribution error isolation).

The model demonstrates robust performance on long-form meetings and short, overlap-rich segments, with notably low Δcp, indicating reliable speaker attribution.

Future work includes streaming SATS, finer-grained timestamp evaluation, and broader multilingual robustness.

02 From the paper

Abstract

Speaker-Attributed, Time-Stamped Transcription (SATS) aims to transcribe what is said and to precisely determine the timing of each speaker, which is particularly valuable for meeting transcription. Existing SATS systems rarely adopt an end-to-end formulation and are further constrained by limited context windows, weak long-range speaker memory, and the inability to output timestamps. To address these limitations, we present MOSS Transcribe Diarize, a unified multimodal large language model that jointly performs Speaker-Attributed, Time-Stamped Transcription in an end-to-end paradigm. Trained on extensive real wild data and equipped with a 128k context window for up to 90-minute inputs, MOSS Transcribe Diarize scales well and generalizes robustly. Across comprehensive evaluations, it outperforms state-of-the-art commercial systems on multiple public and in-house benchmarks.