OpenASR/redimnet2-b6-cn

ReDimNet2-B6 Speaker Embedder (CN-enhanced) · OpenASR

ReDimNet2-B6 speaker embedder for OpenASR diarization — 192-d CN-enhanced embeddings, fully on-device

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

---

✨ Highlights

• 🗣️ OpenASR speaker embedder — the only supported speaker-embedding pack for diarization and Voice ID; required for anonymous speaker labels on any ASR family
• 🧬 192-dim ReDimNet2-B6 — PalabraAI's dimension-reshaping speaker net (12.5M params) with a Chinese-enhanced vb2+vox2+cnc2 training mix
• 🔒 Diarization, not identification — anonymous session-relative labels; embeddings stay local and are discarded after the request unless you explicitly enroll a local profile
• 🎯 Parity-gated packaging — ggml-graph forward pass matches the upstream Python reference at cosine ≥ 0.9999 on held-out fixtures
• 🦀 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 redimnet2-b6-cn:fp16

# 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	redimnet2-b6-cn-fp16.oasr	28 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 ReDimNet2-B6 Speaker Embedder (CN-enhanced)

ReDimNet2-B6 is PalabraAI's 12.5M-parameter speaker-embedding model from the ReDimNet2 family, trained on a VoxBlink2 + VoxCeleb2 + CN-Celeb2 mix so English and Chinese speakers share one embedding space. OpenASR packages the MIT-licensed checkpoint as a local .oasr capability pack and runs it through a ggml graph (not a pure-Rust hand-written forward). This is the only supported speaker-embedding stage for diarization and Voice ID: when the pack is missing, diarize/Voice ID requests fail closed rather than falling back to another embedder. Embeddings are 192-d cosine vectors with a ReDimNet-specific calibration profile.

⚙️ How this pack was made

Converted from PalabraAI/redimnet2 with the OpenASR importer:

openasr model-pack import redimnet2 <src>.safetensors <out>.oasr \
  --package-id redimnet2-b6-cn

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: MIT (source). OpenASR packaging retains the upstream copyright; the only modification is format conversion.

🙏 Acknowledgements

This pack redistributes PalabraAI/redimnet2 checkpoint b6-vb2+vox2+cnc2_v0-lm.pt in OpenASR's .oasr runtime format. Credit for the model architecture, training, and original weights belongs to the upstream PalabraAI / ReDimNet2 authors (paper: ReDimNet2: Scaling Speaker Verification via Time-Pooled Dimension Reshaping). The upstream model is licensed under MIT; OpenASR packaging retains that license and attribution, with the only modification being format conversion for local ggml-graph loading.

🔗 Links

• 🦀 OpenASR — https://github.com/QuintinShaw/openasr
• 🌐 Website — https://openasr.org
• 🤗 Upstream model — PalabraAI/redimnet2