OpenASR/diarizen-large-s80-v2

DiariZen Large-s80-md-v2 · OpenASR

DiariZen Large-s80-md-v2 — optional high-accuracy overlap-aware speaker segmentation for local Voice ID

Speaker-diarization support pack for the OpenASR runtime — pure-Rust inference, no Python at inference time.

---

✨ Highlights

• 🗣️ Overlap-aware local activity — a WavLM Large + Conformer EEND segmenter predicts four local speaker streams in each 16 s window
• 🎯 Qualified native accuracy — OpenASR's fp16 pipeline measured 7.9491% DER on six locked Mandarin meeting excerpts, versus 18.6787% for the MOSS baseline under the same scoring protocol
• 🤝 Shared Voice ID contract — combines with FireRed Stream-VAD, ReDimNet2-B6 and automatic clustering; ASR models consume one normalized speaker timeline
• 🔒 Explicit non-commercial consent — the checkpoint is CC BY-NC 4.0 and is never downloaded or activated without acknowledgement; segmentation-3.0 remains the permissive default
• 🦀 Native fp16 runtime — the .oasr pack runs locally through OpenASR's persistent ggml graph without Python at inference time
• 🦀 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 the pack
openasr pull diarizen-large-s80-v2:fp16 --accept-license

# 3. Diarize any transcription (works with every OpenASR ASR model)
openasr transcribe meeting.wav --model xasr-zh-en --diarize --format srt

📦 Pack

Quant	File (.oasr)	Size
fp16	diarizen-large-s80-v2-fp16.oasr	139 MB

_{Single fp16 build: projection weights ship as fp16; norms/biases and other parity-sensitive tensors stay f32 inside the pack. No extra public quant tiers.}

🧠 About DiariZen Large-s80-md-v2

DiariZen Large-s80-md-v2 is BUT Speech@FIT's overlap-aware speaker-segmentation checkpoint built from a 24-layer WavLM Large encoder and a Conformer EEND head. OpenASR packages the pinned checkpoint as one fp16 .oasr capability pack and uses it as an optional external segmenter in the universal local-file Voice ID pipeline. It predicts recording-local speaker activity; ReDimNet2-B6 still provides clustering and enrolled-person identity. The checkpoint is licensed under CC BY-NC 4.0, so downloading and activating it require explicit non-commercial acknowledgement. OpenASR does not select it merely because Voice ID was enabled; the permissive segmentation-3.0 pack remains the default.

⚙️ How this pack was made

Converted from BUT-FIT/diarizen-wavlm-large-s80-md-v2 with the OpenASR importer:

python3 tooling/diarizen/convert_diarizen.py --checkpoint <pytorch_model.bin> --config <config.toml> --out <diarizen-large-s80-v2-fp16.oasr> --model-id diarizen-large-s80-v2 --quant fp16

The .oasr container is GGUF-backed; projection weights are stored as fp16 while norms/biases and other parity-sensitive tensors remain f32.

⚖️ License

This pack inherits the upstream model's license: CC BY-NC 4.0 (source). OpenASR packaging retains the upstream copyright; the only modification is format conversion.

🙏 Acknowledgements

This pack redistributes the pinned BUT-FIT/diarizen-wavlm-large-s80-md-v2 checkpoint in OpenASR's .oasr runtime format. Credit for the model, architecture, training and original weights belongs to BUT Speech@FIT and the DiariZen authors. The checkpoint is licensed under CC BY-NC 4.0; OpenASR's format conversion does not broaden that license.

🔗 Links

• 🦀 OpenASR — https://github.com/QuintinShaw/openasr
• 🌐 Website — https://openasr.org
• 🤗 Upstream model — BUT-FIT/diarizen-wavlm-large-s80-md-v2