🤗 MOSS-Transcribe-Diarize-0.9B is now open source on Hugging Face. Built with an end-to-end audio-to-structured-transcript paradigm: >0.9B open-source ASR model >Apache license 2.0 >128k long-context transcription >Up to ~90-min audio input >Speaker labels + timestamps in one generation >Multi-speaker diarization for meetings, interruptions, and overlapping voices >Hotword biasing for names,... terms, and domain-specific vocabulary >~100 token/s on NVIDIA RTX 4090, RTF ~0.017 Thank you SGLang vLLM Prince Canuma Lucas Newman for day-0 support! 🚀 Github: Huggingface: API: Live demo: Technical Report： HF Space: AtomGit： SGLang-Omni: vLLM: MLX-audio: Discord:show more

Performance & Benchmarks MOSS-Transcribe-Diarize-0.9B is built for messy, real-world speech: multi-speaker meetings, interruptions, long conversations, and partially overlapping speech. Instead of running ASR, diarization, and alignment as separate steps, the model handles transcription, speaker attribution, and timestamp prediction within a unified generation framework. Trained on million-hour-scale speech data and equipped with a 128k-token context window, MOSS-Transcribe-Diarize-0.9B supports up to ~90 minutes of audio input in a single pass. No slicing, stitching, or complex post-processing is required. It directly generates structured transcripts with speaker labels and timestamps. And it does this without relying on a massive model. The open-source model has only 0.9B parameters. On NVIDIA RTX 4090 single-concurrency setup, it reaches around 100 token/s with an RTF of about 0.017. In practice, a 5–10 minute audio file can be transcribed in under 30 seconds, balancing accuracy, speed, and deployment cost. MOSS-Transcribe-Diarize-0.9B also supports hotword biasing. Users can provide names, company names, product models, project code names, domain terms, and other keywords ahead of time to improve recognition of important terms in specific business scenarios. This makes the model suitable not only for general meeting transcription, but also for enterprise meeting notes, customer communication records, call-center QA, interview transcription, and industry-specific content workflows. On AISHELL-4, MOSS-Transcribe-Diarize-0.9B achieves the best results across three key metrics: CER, cpCER, and Δcp. Its cpCER drops to 14.98%, reducing error by more than 40% compared with mainstream existing models. Its Δcp is only 0.79, indicating stable and accurate speaker attribution in long multi-speaker conversations. In other words, MOSS-Transcribe-Diarize-0.9B does not just transcribe more accurately. It also keeps track of who is speaking across the whole conversation.

End-to-End: Rethinking Multi-Speaker Transcription Compared with cascaded systems such as WhisperX, which typically rely on separate ASR, diarization, and alignment modules, the core idea behind MOSS-Transcribe-Diarize-0.9B is to model speech transcription, speaker attribution, and timestamp prediction as a single autoregressive generation task. In other words, the model no longer solves “what was said,” “who said it,” and “when it was said” as separate problems. Instead, it handles all three within one unified generation process. In terms of architecture, MOSS-Transcribe-Diarize-0.9B uses a multimodal design with a Whisper-Medium audio encoder and a Qwen3-0.6B-style causal decoder. The audio encoder extracts acoustic information from long-form, multi-speaker speech, while the language model directly generates structured transcripts with timestamps and speaker labels. [0.11] [S01] Good morning! [1.03] [1.11] [S02] Morning, guys! [1.34] This design brings three key advantages: 1. A truly end-to-end experience: no separate ASR, speaker diarization, or alignment modules are required. 2. Text-level time modeling: time information is discretized into generatable tokens and modeled together with text, improving timestamp stability in long-context scenarios. 3. Global consistency modeling: speaker identity, transcript content, and timing information are jointly optimized within the same context, enabling more stable long-range modeling.

Multi-speaker diarization

Multilingual transcription

Try the PRO model live:

@huggingface Wow... I just tried it with pytoch. Even though the clip was 5 minutes of Japanese, the transcription was surprisingly good with very few errors, considering the model size. Can you guys please share when it will be available in CrispASR?

@huggingface thank you for trying our model! let me call our tech bro to answer @YangGao07

@huggingface @YangGao07 Thanks <3

@huggingface 🔥🔥🔥

@huggingface @Presidentlin

@huggingface I am most impressed with the AI they used to generate the video.

@ClementDelangue @huggingface 这功能看起来挺实用的，期待试试效果！

@huggingface 90分钟音频一次出标签和时间戳，这个挺省后处理的

@huggingface 期待你的测评！🤝🤝🤝

@huggingface Looks promising.

@huggingface Thank you!

@huggingface 128kのロングコンテキストで丸ごと文字起こしできるの、地味にありがたいですね。自分も相場解説の長尺動画を後で見返すのだるくて溜めがちなので、これで一気に構造化してくれるなら普通に使いたいです😊

@huggingface Its very very good! Great work!

@huggingface Glad to hear that 😊

@huggingface how many speakers. What languages are supported? is it possible to get speakers voice-stamps so that we could easily identify them in next conversation?

@huggingface unlimited speakers. speakers has voice stamps yes, it's gonna look like this for the specific languages will need to check with the team, theoretically most of the popular languages are all covered

@huggingface How the hell did you guys create this promo video?

@huggingface To the Moon 🌕

@huggingface I'm impressed you got 128k context and diarization in one model. 0.017 RTF on a 4090 is a practical speed for meeting logs. Open source will let many teams build their own pipelines

@huggingface best！！！！😍😍😍

@huggingface 😍😍😍😍😍😍😍😍😍

@huggingface Structured transcription is more useful than raw ASR for agents. Speaker labels, timestamps, and long context turn audio into something a meeting/workflow agent can cite, segment, and hand off. The product primitive is verifiable evidence, not just text.

@huggingface Impressive release, though real-world accuracy on noisy audio and overlapping speakers will matter more than specs. Looking forward to independent benchmarks.

@huggingface Glad to hear 🥰

@huggingface 90 min audio input with speaker diarization all open source is a solid drop for meeting transcription.

@huggingface 👀😍

@huggingface Impressive

@huggingface Impressive open-source release. ASR, diarization, and timestamps in one pipeline is a big deal.

@huggingface great upgrades MTD!

@huggingface yesss

@huggingface It works amazing as end2end, but what if i want to extract speaker embeddings/fingerprint?

@victormustar @huggingface Smol and nice !!