OpenMOSS has released MOSS-Transcribe-Diarize, a 0.9B parameter end-to-end audio understanding model designed for long-form multi-speaker transcription, diarization, timestamps, and acoustic event awareness.
The model jointly performs speech transcription and speaker diarization in a single pass, eliminating the need to stitch together separate ASR and diarization systems. It generates compact, time-aligned transcripts with anonymous speaker labels such as [S01] and [S02]. Built on a Qwen3-0.6B style causal decoder and a Whisper-Medium encoder, it supports custom transcription instructions, hotwords, and acoustic event annotations.
This approach provides a richer view of audio content for downstream systems by identifying who spoke and when within meetings, calls, podcasts, and other long recordings.