OpenASR/moss-transcribe-diarize

<div align="center">

MOSS-Transcribe-Diarize · OpenASR

Joint transcription and speaker diarization in a single 0.9B-parameter model — know who said what, and when

   

Native speech-to-text in the [OpenASR](https://github.com/QuintinShaw/openasr) runtime — engineered for peak performance on CPU & GPU, no Python at inference time.

</div>

✨ Highlights

• —🎯 Transcription + speaker identification in one pass — a single inference run produces both the transcript and per-speaker timestamps, no separate pipeline needed
• —📊 Verified accuracy: 2.52% CER (Chinese), 2.23% WER (English) — benchmarked by OpenASR on frozen evaluation datasets, not upstream-reported numbers
• —🌐 Chinese and English first, 15+ additional languages — optimized for Mandarin and English with broad multilingual coverage built in
• —📦 Three quantization tiers: fp16 / q8_0 / q4_k — delivered in OpenASR's native .oasr format for CPU inference, choose the precision-to-size tradeoff that fits your hardware
• —🦀 Native in OpenASR — .oasr packs run with no Python at inference, engineered for peak performance on CPU & GPU

🚀 Quickstart

# 1. Install the OpenASR CLI  ·  https://openasr.org
# 2. Pull a build (pick a quant — see the table below)
openasr pull moss-transcribe-diarize:q8

# 3. Transcribe
openasr transcribe audio.wav --model moss-transcribe-diarize

All builds for this model:

openasr pull moss-transcribe-diarize:fp16
openasr pull moss-transcribe-diarize:q8
openasr pull moss-transcribe-diarize:q4

📦 Available builds

_{RTF = real-time factor on the fixed 11s JFK clip (lower is faster); RAM peak measured per pack in an isolated subprocess. JFK ΔWER compares each quantized build's JFK transcript to this model's fp16 JFK transcript, so it measures quantization drift rather than absolute recognition accuracy. q8_0 is the recommended default — near-reference quality at a fraction of the footprint.}

🧠 About MOSS-Transcribe-Diarize

MOSS-Transcribe-Diarize is a 0.9B-parameter joint speech-recognition and speaker-diarization model from the OpenMOSS team. Its architecture pairs a Whisper-Medium-scale audio encoder with a lightweight MLP/LayerNorm adapter feeding into a Qwen3-0.6B-scale decoder, enabling the model to transcribe speech and attribute each segment to its speaker with start/end timestamps in a single forward pass. Language coverage centers on Chinese and English, with support for 15 additional languages. OpenASR distributes this model in three quantization tiers -- fp16, q80, and q4k -- packaged in the native .oasr runtime format for CPU-based local inference.

⚙️ How these packs were made

Converted from OpenMOSS-Team/MOSS-Transcribe-Diarize with the OpenASR importer:

openasr model-pack import moss <src> <out>.oasr \
  --package-id moss-transcribe-diarize --quantization {fp16,q8-0,q4-k}

The .oasr container is GGUF-backed; packs use zero-copy mmap weight binding and graph buffer reuse to keep peak memory low.

⚖️ License

These packs inherit the upstream model's license: Apache-2.0 (source). OpenASR packaging retains the upstream copyright and NOTICE; the only modifications are format conversion and quantization.

🙏 Acknowledgements

This pack is a redistribution of MOSS-Transcribe-Diarize, created and released by OpenMOSS-Team (OpenMOSS-Team/MOSS-Transcribe-Diarize) under the Apache License 2.0. OpenASR performs format conversion, quantization, runtime validation, and local-inference adaptation only; all model weights and training are the work of the original authors. Thank you to the OpenMOSS team for releasing their work openly.

🔗 Links

• —🦀 OpenASR — <https://github.com/QuintinShaw/openasr>
• —🌐 Website — <https://openasr.org>
• —🤗 Upstream model — OpenMOSS-Team/MOSS-Transcribe-Diarize