areffarhadi/w2v-bert-VoxLingua

VoxLingua LID — WPT + MHFA (Odyssey 2026)

Language identification checkpoint trained on VoxLingua107, evaluated on dev_vox.

Related paper: Subtract to Clean, Add to Enrich: Dual-Path Disentanglement for Speaker and Language Recognition (Odyssey 2026).

Author: Aref Farhadipour

Files

File	Description
model.pt	WPT + W2V-BERT-2.0 + MHFA classifier (107 languages)
args.json	Training hyperparameters

• Backbone: facebook/w2v-bert-2.0
• Prompt tokens: 15 regular + 8 wavelet per layer
• Clip length: 96000 samples (6 s @ 16 kHz)

Reported performance: see the paper.

Download

huggingface-cli download areffarhadi/w2v-bert-VoxLingua --local-dir ./checkpoints/voxlingua_lid

Code

GitHub: areffarhadi/ASV-LID

Eval

Dataset layout: dev_vox/<lang>/*.wav

bash scripts/eval_voxlingua_lid.sh 0

Citation

@inproceedings{farhadipour2026subtract,
  title={Subtract to Clean, Add to Enrich: Dual-Path Disentanglement for Speaker and Language Recognition},
  author={Farhadipour, Aref},
  booktitle={Proc. Odyssey},
  year={2026}
}