ASR and diarization, finally shoved into a single 0.9B model

What it does

MOSS-Transcribe-Diarize 0.9B ingests raw audio and emits timestamped transcripts where each utterance is prefixed with a speaker tag like [S01]. It handles 50+ languages and is built for messy real-world recordings—meetings, podcasts, lectures—rather than clean single-speaker clips. The model also claims optional acoustic event awareness, though the README does not quantify that capability.

The interesting bit

Instead of chaining a separate diarization engine after an ASR model, it fuses a Whisper-Medium audio encoder with a Qwen3-style text decoder through a 4× temporal merge and MLP adaptor, then generates speaker labels and text autoregressively in one shot. That architectural choice keeps the whole system under a billion parameters while still placing first or second on the provided AISHELL-4, Alimeeting, Podcast, and Movies benchmarks against much larger closed APIs.

Key highlights

• End-to-end output: emits compact [start][Sxx]text[end] segments without post-processing alignment.
• 50+ language support out of the box.
• Serves through an OpenAI-compatible /v1/audio/transcriptions endpoint via SGLang Omni or vLLM.
• Includes a browser-based subtitle Web UI (now in English and Simplified Chinese).
• A stronger “Pro” variant exists, but is API-only; the 0.9B weights are fully open.

Caveats

• The README warns that the eager attention backend is a last resort: its memory usage scales quadratically with audio length and will OOM on long recordings.
• SGLang Omni serving currently targets CUDA 13; CUDA 12 users must fall back to vLLM.
• Acoustic event awareness is mentioned as a feature, yet no evaluation metrics for it appear in the provided sources.

Verdict Worth a look if you need offline, speaker-aware transcription for long-form content and want to avoid gluing together Whisper and PyAnnote. Skip it if you only need single-speaker ASR or if you are locked to CUDA 12 and need SGLang-specific features.