witness-redimnet2-b6-mlx

ReDimNet2-B6-lm speaker embedder (192-d, raw-16kHz-waveform in), converted to MLX safetensors for witness's on-device speaker diarization.

Upstream attribution

• Model: ReDimNet2-B6-lm, checkpoint b6-vox2-lm.pt (vox2, lm).
• Authors / source: PalabraAI — PalabraAI/redimnet2 (arXiv 2603.11841).
• License: MIT (inherited from PalabraAI/redimnet2).

What's in this repo

• model.safetensors — weights converted verbatim from the PyTorch state-dict (num_batches_tracked dropped, f32) by .research/diarization/gen_redimnet2_embed_fixture.py.
• NOTES.md, TENSOR_SHAPES.txt — the conversion reference + full tensor shape manifest the MLX C++ loader keys on.

192-d embedding, not L2-normalized by the model (the witness crate L2-normalizes after). Parity vs the PyTorch reference is 1 − cos ≤ 5e-7.

---

Converted to MLX for witness, an open-source Rust toolkit for on-device system capture on macOS. Generated by .research/diarization/publish_weights.sh.