MOSS-Transcribe

Meeting transcription normally chains ASR, voice activity detection, and a separate diarization model; MOSS-Transcribe-Diarize folds the whole job into one end-to-end model that ingests long multi-speaker recordings and emits a single stream of speaker-labeled, timestamped text with acoustic events included. The 0.9B open release pairs a Whisper-Medium-shaped audio encoder with a Qwen3-0.6B-style decoder, covers more than 50 languages, and took first place in the second MLC-SLM challenge at Interspeech 2026. OpenMOSS ships a Python package, serving recipes for vLLM and SGLang, a local subtitle web app exporting JSON, SRT, and ASS files, and a fine-tuning framework, all Apache-2.0 with weights on Hugging Face; a larger Pro model sits behind the same interface as a preview. Audio is processed at 16kHz in 30-second chunks and inference is GPU-oriented, with throughput quoted on H100 but a footprint small enough for consumer cards. Released July 2026, it is the strongest open answer to who-said-what transcription.