EfficientTDNN

This repository provides all the necessary tools to perform speaker verification with a NAS alternative, named as EfficientTDNN. The system can be used to extract speaker embeddings with different model size. It is trained on Voxceleb2 training data using data augmentation. The model performance on Voxceleb1-test set(Cleaned)/Vox1-O are reported as follows.

Supernet Stage	Subnet	MACs (3s)	Params	EER(%)	minDCF
depth	Base	1.45G	5.79M	0.94	0.089
width 1	Mobile	570.98M	2.42M	1.41	0.124
width 2	Small	204.07M	899.20K	2.20	0.219

The details of three subnets are:

Base: (3, [512, 512, 512, 512], [5, 3, 3, 3], 1536)
Mobile: (3, [384, 256, 256, 256], [5, 3, 3, 3], 768)
Small: (2, [256, 256, 256], [3, 3, 3], 400)

Compute your speaker embeddings

import torch
from sugar.models import WrappedModel
wav_input_16khz = torch.randn(1,10000).cuda()

repo_id = "mechanicalsea/efficient-tdnn"
supernet_filename = "depth/depth.torchparams"
subnet_filename = "depth/depth.ecapa-tdnn.3.512.512.512.512.5.3.3.3.1536.bn.tar"
subnet, info = WrappedModel.from_pretrained(repo_id=repo_id, supernet_filename=supernet_filename, subnet_filename=subnet_filename)
subnet = subnet.cuda()
subnet = subnet.eval()

embedding = subnet(wav_input_16khz)

Inference on GPU

To perform inference on the GPU, add subnet = subnet.to(device) after calling the from_pretrained method.

Model Description

Models are listed as follows.

Dynamic Kernel: The model enables various kernel sizes in {1,3,5}, kernel/kernel.torchparams.
Dynamic Depth: The model enables additional various depth in {2,3,4} based on Dynamic Kernel version, depth/depth.torchparams.
Dynamic Width 1: The model enable additional various width in [0.5, 1.0] based on Dynamic Depth version, width1/width1.torchparams.
Dynamic Width 2: The model enable additional various width in [0.25, 0.5] based on Dynamic Width 1 version, width2/width2.torchparams.

Furthermore, some subnets are given in the form of the weights of batchnorm corresponding to their trained supernets as follows.

Dynamic Kernel
1. kernel/kernel.max.bn.tar
2. kernel/kernel.Kmin.bn.tar
Dynamic Depth
1. depth/depth.max.bn.tar
2. depth/depth.Kmin.bn.tar
3. depth/depth.Dmin.bn.tar
4. depth/depth.3.512.5.5.3.3.1536.bn.tar
5. depth/depth.ecapa-tdnn.3.512.512.512.512.5.3.3.3.1536.bn.tar
Dynamic Width 1
1. width1/width1.torchparams
2. width1/width1.max.bn.tar
3. width1/width1.Kmin.bn.tar
4. width1/width1.Dmin.bn.tar
5. width1/width1.C1min.bn.tar
6. width1/width1.3.383.256.256.256.5.3.3.3.768.bn.tar
Dynamic Width 2
1. width2/width2.max.bn.tar
2. width2/width2.Kmin.bn.tar
3. width2/width2.Dmin.bn.tar
4. width2/width2.C1min.bn.tar
5. width2/width2.C2min.bn.tar
6. width2/width2.3.384.3.1152.bn.tar
7. width2/width2.3.256.256.384.384.1.3.5.3.1152.bn.tar
8. width2/width2.2.256.256.256.3.3.3.400.bn.tar

The tag is described as follows.

max: (4, [512, 512, 512, 512, 512], [5, 5, 5, 5, 5], 1536)
Kmin: (4, [512, 512, 512, 512, 512], [1, 1, 1, 1, 1], 1536)
Dmin: (2, [512, 512, 512], [1, 1, 1], 1536)
C1min: (2, [256, 256, 256], [1, 1, 1], 768)
C2min: (2, [128, 128, 128], [1, 1, 1], 384)

More details about EfficentTDNN can be found in the paper EfficientTDNN.

Citing EfficientTDNN

Please, cite EfficientTDNN if you use it for your research or business.

@article{wr-efficienttdnn-2022,
  author={Wang, Rui and Wei, Zhihua and Duan, Haoran and Ji, Shouling and Long, Yang and Hong, Zhen},
  journal={IEEE/ACM Transactions on Audio, Speech, and Language Processing}, 
  title={EfficientTDNN: Efficient Architecture Search for Speaker Recognition}, 
  year={2022},
  volume={30},
  number={},
  pages={2267-2279},
  doi={10.1109/TASLP.2022.3182856}}