Sugar

Efficient Speech Processing Tookit for Automatic Speaker Recognition

| HuggingFace |

The authors' PyTorch implementation and pretrained models of EfficientTDNN.

• 17 June 2022: EfficientTDNN are published in IEEE/ACM Transactions on Audio, Speech, and Language Processing.

What's New

• EfficientTDNN: Efficient Architecture Search for Speaker Recognition [arXiv] [IEEE/ACM TASLP]

Models and Checkpoints

Model	Training Dataset	Link
EfficientTDNN	VoxCeleb2	HuggingFace: mechanicalsea/efficient-tdnn

Requirements and Installation

• PyTorch version >= 1.7.1
• Python version >= 3.7.9
• To install sugar:

git clone https://github.com/mechanicalsea/sugar.git
cd sugar
pip install --editable .

# [Option] Requrements Packages: numpy, scipy, pandas, matplotlib, pyyaml, thop, geatpy, torch, torchvision, torchaudio, scikit-learn

We provide pre-trained models for extracting speaker embeddings via huggingface.

Tutorials

• EfficientTDNN
  • the evalution of a subnet
  • evaluation of different subnets from a given supernet
  • search on the trained supernet

EfficientTDNN Training Scripts

An example of training the TDNN supernet is given as follows.

task=''            # training stage, e.g., supernet, supernet_kernel, supernet_kernel_depth, ...
cycle_step=34120   # a half cycle of learning rate scheduler, e.g., GPU 1: 68248, GPU 2: 34120, GPU 4: 17056
second=48000       # largest stage: 32000, other stages: 48000
initial_weights='' # required to these stages except for the largest stage
logdir=''
python scripts/train_vox2_veri.py --distributed --augment \
      --task supernet_width1 --epochs 64 --cycle-step ${cycle_step} --second ${second} \
      --initial-weights ${initial_weights} \
      --logdir ${logdir}

Load Pre-Trained Models for Inference

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)

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}}

Contact Information

For help or issues using EfficientTDNN models, please submit a GitHub issue.

For other communications related to EfficientTDNN, please contact Rui Wang (rwang@tongji.edu.cn).