0.9B open end-to-end ASR: one pass does transcription + speaker diarization + timestamps + optional acoustic events. 50+ languages. Apache 2.0.
Output looks like [0.48][S01]Welcome everyone[1.66]. Built for long messy meetings/podcasts/lectures instead of stitching a separate diarizer onto Whisper.
Won 1st in the 2nd MLC-SLM Challenge (INTERSPEECH 2026, 14 languages). Serve via SGLang Omni, vLLM, or FunASR. A stronger “Pro” variant is playground-only.
