FAIR-GATE ECAPA VoxCeleb r=0.05

This repository contains the final FAIR-GATE checkpoint for gender-fair automatic speaker verification with an ECAPA-TDNN backbone.

Checkpoint

  • Checkpoint file: fairgate_ecapa_voxceleb_r005_epoch30.pt
  • Backbone: ECAPA-TDNN
  • Final configuration: r=0.05
  • Selection checkpoint: epoch 30

Configuration

--css_target_ratio 0.05
--lambda_css_cap 0.005
--lambda_css_sat 0.001
--lambda_gender_s 0.05
--lambda_gender_adv 0.002
--grl_warmup_epochs 8
--adv_warmup_epochs 8

Intended use

This checkpoint is intended for research on automatic speaker verification and fairness analysis.

Code

The evaluation code is available at:

https://github.com/<your-github-username>/fairgate-asv

Evaluation

Clone the code repository first, then run evaluation from the code repository directory:

cd FAIR-GATE-ASV

EVAL_PATH=/path/to/VoxCeleb1 \
GENDER_MAP=/path/to/gender_map_eval.json \
VOX1_O=/path/to/vox1-O-abs.txt \
VOX1_E=/path/to/vox1-E-abs.txt \
VOX1_H=/path/to/vox1-H-abs.txt \
SAVE_PATH=exps/eval_no_asnorm \
N_CLASS=5994 \
CUDA_VISIBLE_DEVICES=0 \
bash scripts/evaluation/evaluate_vox1_no_asnorm.sh /path/to/fairgate_ecapa_voxceleb_r005_epoch30.pt

Limitations

This model was developed for research purposes. Performance and fairness behavior may vary across datasets, demographic distributions, recording conditions, and operating thresholds.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support