ASV-Subtools Recipe

⚠️ This project is currently under development.

⚠️ The training settings in this recipe are different from those in the D-TDNN paper:

	Recipe	Paper
Train on	VoxCeleb2 dev + aug * 4	VoxCeleb1 dev + VoxCeleb2
VAD	✔️	✖️
# Feature	81	23
# Frame	200	200-400
Nonlinearity	ReLU	PReLU
Momentum (last BN)	0.5	0.1
Margin	0.25	0.4
Scaling factor	32	64
Batch size	256	128
Momentum (SGD)	0.9	0.95
Initial LR	0.02	0.01
LR scheduler	ReduceLROnPlateau	Fixed

⚠️ Up to now the training settings in the D-TDNN paper achieve better performance.

Usage

cd $kaldi_root/egs
mkdir asv && cd asv
git clone https://github.com/yuyq96/asv-subtools.git subtools
cd subtools/recipe
git clone https://github.com/yuyq96/subtools-recipe-voxceleb.git voxceleb-dtdnn
cd ../..
subtools/recipe/runVoxceleb.sh