causal/redimnet2-b3-vb2-vox2-cnc2-lm-mlx

ReDimNet2 B3 (VoxBlink2 + VoxCeleb2 + CN-Celeb2, lm) for MLX

MLX conversion of the official ReDimNet2 b3-vb2+vox2+cnc2_v0-lm.pt checkpoint (large-margin fine-tuned (recommended)). It produces 192-dimensional speaker embeddings from 16 kHz mono audio.

Usage

uv add git+https://github.com/avra-m3/speaker-embedding-mlx

import mlx.core as mx
from redimnet2_mlx import load_model

model = load_model("causal/redimnet2-b3-vb2-vox2-cnc2-lm-mlx")
emb = model(mx.array(wav_16k_float32)[None])  # (1, 192)

Parity

Embeddings match the PyTorch reference to 3.2e-06 relative error (worst case, float32, MLX CPU backend). See the parity log.

Training data terms

This checkpoint was trained on VoxBlink2 and CN-Celeb in addition to VoxCeleb2. The model code and weights are MIT licensed upstream, but those datasets have their own terms of use, which may restrict use to research. Check them before any commercial use.

License and citation

MIT, following the upstream ReDimNet2 release (Copyright (c) 2026 Palabra.ai). See third-party notices.

@inproceedings{redimnet2_2026,
  title={ReDimNet2: Scaling Speaker Verification via Time-Pooled Dimension Reshaping},
  author={Yakovlev, Ivan and Okhotnikov, Anton},
  booktitle={Proceedings of Interspeech 2026},
  year={2026},
  eprint={2603.11841},
  archivePrefix={arXiv},
}