PyTorch implementation of Densely Connected Time Delay Neural Network

Overview

Densely Connected Time Delay Neural Network

PyTorch implementation of Densely Connected Time Delay Neural Network (D-TDNN) in our paper "Densely Connected Time Delay Neural Network for Speaker Verification" (INTERSPEECH 2020).

What's New ⚠️

  • [2021-02-14] We add an impl option in TimeDelay, now you can choose:

    • 'conv': implement TDNN by F.conv1d.
    • 'linear': implement TDNN by F.unfold and F.linear.

    Check this commit for more information. Note the pre-trained models of 'conv' have not been uploaded yet.

  • [2021-02-04] TDNN (default implementation) in this repo is slower than nn.Conv1d, but we adopted it because:

    • TDNN in this repo was also used to create F-TDNN models that are not perfectly supported by nn.Conv1d (asymmetric paddings).
    • nn.Conv1d(dilation>1, bias=True) is slow in training.

    However, we do not use F-TDNN here, and we always set bias=False in D-TDNN. So, we are considering uploading a new version of TDNN soon (2021-02-14 updated).

  • [2021-02-01] Our new paper is accepted by ICASSP 2021.

    Y.-Q. Yu, S. Zheng, H. Suo, Y. Lei, and W.-J. Li, "CAM: Context-Aware Masking for Robust Speaker Verification"

    CAM outperforms statistics-and-selection (SS) in terms of speed and accuracy.

Pretrained Models

We provide the pretrained models which can be used in many tasks such as:

  • Speaker Verification
  • Speaker-Dependent Speech Separation
  • Multi-Speaker Text-to-Speech
  • Voice Conversion

D-TDNN & D-TDNN-SS

Usage

Data preparation

You can either use Kaldi toolkit:

  • Download VoxCeleb1 test set and unzip it.
  • Place prepare_voxceleb1_test.sh under $kaldi_root/egs/voxceleb/v2 and change the $datadir and $voxceleb1_root in it.
  • Run chmod +x prepare_voxceleb1_test.sh && ./prepare_voxceleb1_test.sh to generate 30-dim MFCCs.
  • Place the trials under $datadir/test_no_sil.

Or checkout the kaldifeat branch if you do not want to install Kaldi.

Test

  • Download the pretrained D-TDNN model and run:
python evaluate.py --root $datadir/test_no_sil --model D-TDNN --checkpoint dtdnn.pth --device cuda

Evaluation

VoxCeleb1-O

Model Emb. Params (M) Loss Backend EER (%) DCF_0.01 DCF_0.001
TDNN 512 4.2 Softmax PLDA 2.34 0.28 0.38
E-TDNN 512 6.1 Softmax PLDA 2.08 0.26 0.41
F-TDNN 512 12.4 Softmax PLDA 1.89 0.21 0.29
D-TDNN 512 2.8 Softmax Cosine 1.81 0.20 0.28
D-TDNN-SS (0) 512 3.0 Softmax Cosine 1.55 0.20 0.30
D-TDNN-SS 512 3.5 Softmax Cosine 1.41 0.19 0.24
D-TDNN-SS 128 3.1 AAM-Softmax Cosine 1.22 0.13 0.20

Citation

If you find D-TDNN helps your research, please cite

@inproceedings{DBLP:conf/interspeech/YuL20,
  author    = {Ya-Qi Yu and
               Wu-Jun Li},
  title     = {Densely Connected Time Delay Neural Network for Speaker Verification},
  booktitle = {Annual Conference of the International Speech Communication Association (INTERSPEECH)},
  pages     = {921--925},
  year      = {2020}
}

Revision of the Paper ⚠️

References:

[16] X. Li, W. Wang, X. Hu, and J. Yang, "Selective Kernel Networks," in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 510-519.

Comments
  • size mismatch while loading pre-trained weights

    size mismatch while loading pre-trained weights

    RuntimeError: Error(s) in loading state_dict for DTDNN: Missing key(s) in state_dict: "xvector.tdnn.linear.bias", "xvector.dense.linear.bias". size mismatch for xvector.tdnn.linear.weight: copying a param with shape torch.Size([128, 30, 5]) from checkpoint, the shape in current model is torch.Size([128, 150]). size mismatch for xvector.block1.tdnnd1.linear1.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([128, 128]). size mismatch for xvector.block1.tdnnd1.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.block1.tdnnd2.linear1.weight: copying a param with shape torch.Size([128, 192, 1]) from checkpoint, the shape in current model is torch.Size([128, 192]). size mismatch for xvector.block1.tdnnd2.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.block1.tdnnd3.linear1.weight: copying a param with shape torch.Size([128, 256, 1]) from checkpoint, the shape in current model is torch.Size([128, 256]). size mismatch for xvector.block1.tdnnd3.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.block1.tdnnd4.linear1.weight: copying a param with shape torch.Size([128, 320, 1]) from checkpoint, the shape in current model is torch.Size([128, 320]). size mismatch for xvector.block1.tdnnd4.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.block1.tdnnd5.linear1.weight: copying a param with shape torch.Size([128, 384, 1]) from checkpoint, the shape in current model is torch.Size([128, 384]). size mismatch for xvector.block1.tdnnd5.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.block1.tdnnd6.linear1.weight: copying a param with shape torch.Size([128, 448, 1]) from checkpoint, the shape in current model is torch.Size([128, 448]). size mismatch for xvector.block1.tdnnd6.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.transit1.linear.weight: copying a param with shape torch.Size([256, 512, 1]) from checkpoint, the shape in current model is torch.Size([256, 512]). size mismatch for xvector.block2.tdnnd1.linear1.weight: copying a param with shape torch.Size([128, 256, 1]) from checkpoint, the shape in current model is torch.Size([128, 256]). size mismatch for xvector.block2.tdnnd1.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.block2.tdnnd2.linear1.weight: copying a param with shape torch.Size([128, 320, 1]) from checkpoint, the shape in current model is torch.Size([128, 320]). size mismatch for xvector.block2.tdnnd2.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.block2.tdnnd3.linear1.weight: copying a param with shape torch.Size([128, 384, 1]) from checkpoint, the shape in current model is torch.Size([128, 384]). size mismatch for xvector.block2.tdnnd3.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.block2.tdnnd4.linear1.weight: copying a param with shape torch.Size([128, 448, 1]) from checkpoint, the shape in current model is torch.Size([128, 448]). size mismatch for xvector.block2.tdnnd4.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.block2.tdnnd5.linear1.weight: copying a param with shape torch.Size([128, 512, 1]) from checkpoint, the shape in current model is torch.Size([128, 512]). size mismatch for xvector.block2.tdnnd5.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.block2.tdnnd6.linear1.weight: copying a param with shape torch.Size([128, 576, 1]) from checkpoint, the shape in current model is torch.Size([128, 576]). size mismatch for xvector.block2.tdnnd6.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.block2.tdnnd7.linear1.weight: copying a param with shape torch.Size([128, 640, 1]) from checkpoint, the shape in current model is torch.Size([128, 640]). size mismatch for xvector.block2.tdnnd7.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.block2.tdnnd8.linear1.weight: copying a param with shape torch.Size([128, 704, 1]) from checkpoint, the shape in current model is torch.Size([128, 704]). size mismatch for xvector.block2.tdnnd8.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.block2.tdnnd9.linear1.weight: copying a param with shape torch.Size([128, 768, 1]) from checkpoint, the shape in current model is torch.Size([128, 768]). size mismatch for xvector.block2.tdnnd9.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.block2.tdnnd10.linear1.weight: copying a param with shape torch.Size([128, 832, 1]) from checkpoint, the shape in current model is torch.Size([128, 832]). size mismatch for xvector.block2.tdnnd10.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.block2.tdnnd11.linear1.weight: copying a param with shape torch.Size([128, 896, 1]) from checkpoint, the shape in current model is torch.Size([128, 896]). size mismatch for xvector.block2.tdnnd11.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.block2.tdnnd12.linear1.weight: copying a param with shape torch.Size([128, 960, 1]) from checkpoint, the shape in current model is torch.Size([128, 960]). size mismatch for xvector.block2.tdnnd12.linear2.weight: copying a param with shape torch.Size([64, 128, 3]) from checkpoint, the shape in current model is torch.Size([64, 384]). size mismatch for xvector.transit2.linear.weight: copying a param with shape torch.Size([512, 1024, 1]) from checkpoint, the shape in current model is torch.Size([512, 1024]). size mismatch for xvector.dense.linear.weight: copying a param with shape torch.Size([512, 1024, 1]) from checkpoint, the shape in current model is torch.Size([512, 1024]).

    opened by zabir-nabil 3
  • 实验细节的疑问

    实验细节的疑问

    您好: 我想教下您的论文中,实验的实现细节: 1.实验数据:我看很多其他论文都是使用voxceleb2 dev 5994说话人作为训练集(或者voxceleb dev+voxceleb2 dev,1211+5994说话人),您有只在这部分说话人上的实验结果吗?方便透露下嘛?

    2.PLDA和Cosine Similarity:您这里实验比较这两个的EER在TDNN中是提取的是倒数第二层(分类器前一层)还是第三层(xvector)的输出啊?因为我在论文中又看到,这两个不同层embedding对不同方法性能有差异,倒数第二层的cosine方法可能会更好一些。

    Thanks!🙏

    opened by Wenhao-Yang 1
  • questions about model training

    questions about model training

    hello, yuyq96, Thank you so much for the great work you've shared. I learned that D-TDNNSS mini-batch setting 128 from D-TDNN paper. But this model is too large to train on single gpu. Could you tell me how you train it? Using nn.Parallel or DDP? Looking forward to you reply

    opened by forwiat 2
  • the difference between kaldifeat-kaldi and kaldifeat-python?

    the difference between kaldifeat-kaldi and kaldifeat-python?

    May I ask you the numerical difference between kaldifeat by kaldi implementation and kaldifeat by your python implementation? I have compared the two computed features, and I find it has some difference. I wonder that the experiment results showed in D-TDNN master and D-TDNN-kaldifeat branch is absolutely the same.

    Thanks~

    opened by mezhou 4
  • 针对论文的一些疑问

    针对论文的一些疑问

    您好,我觉得您的工作-DTDNN,在参数比较少的情况下获得了较ETDNN,FTDNN更好的结果,我认为这非常有意义。但是我对论文的实验存在两处疑惑: 1、论文中Table5中,基于softmax训练的D-TDNN模型Cosine的结果好于PLDA,在上面的TDNN,ETDNN,FTDNN的结果不一致(均是PLDA好于Cosine),请问这是什么原因导致的? 2、对于null branch,能稍微解释一下吗?

    opened by xuanjihe 10
Releases(trials)
Owner
Ya-Qi Yu
Machine Learning
Ya-Qi Yu
Easy way to add GoogleMaps to Flask applications. maintainer: @getcake

Flask Google Maps Easy to use Google Maps in your Flask application requires Jinja Flask A google api key get here Contribute To contribute with the p

Flask Extensions 611 Dec 05, 2022
Airbus Ship Detection Challenge

Airbus Ship Detection Challenge This is an open solution to the Airbus Ship Detection Challenge. Our goals We are building entirely open solution to t

minerva.ml 55 Nov 29, 2022
It helps user to learn Pick-up lines and share if he has a better one

Pick-up-Lines-Generator(Open Source) It helps user to learn Pick-up lines Share and Add one or many to the DataBase Unique SQLite DataBase AI Undercon

knock_nott 0 May 04, 2022
A GPU-optional modular synthesizer in pytorch, 16200x faster than realtime, for audio ML researchers.

torchsynth The fastest synth in the universe. Introduction torchsynth is based upon traditional modular synthesis written in pytorch. It is GPU-option

torchsynth 229 Jan 02, 2023
Official repository for the paper "Self-Supervised Models are Continual Learners" (CVPR 2022)

Self-Supervised Models are Continual Learners This is the official repository for the paper: Self-Supervised Models are Continual Learners Enrico Fini

Enrico Fini 73 Dec 18, 2022
Self-Supervised Monocular DepthEstimation with Internal Feature Fusion(arXiv), BMVC2021

DIFFNet This repo is for Self-Supervised Monocular Depth Estimation with Internal Feature Fusion(arXiv), BMVC2021 A new backbone for self-supervised d

Hang 94 Dec 25, 2022
Deep Text Search is an AI-powered multilingual text search and recommendation engine with state-of-the-art transformer-based multilingual text embedding (50+ languages).

Deep Text Search - AI Based Text Search & Recommendation System Deep Text Search is an AI-powered multilingual text search and recommendation engine w

19 Sep 29, 2022
Callable PyTrees and filtered JIT/grad transformations => neural networks in JAX.

Equinox Callable PyTrees and filtered JIT/grad transformations = neural networks in JAX Equinox brings more power to your model building in JAX. Repr

Patrick Kidger 909 Dec 30, 2022
[UNMAINTAINED] Automated machine learning for analytics & production

auto_ml Automated machine learning for production and analytics Installation pip install auto_ml Getting started from auto_ml import Predictor from au

Preston Parry 1.6k Jan 02, 2023
Unsupervised Domain Adaptation for Nighttime Aerial Tracking (CVPR2022)

Unsupervised Domain Adaptation for Nighttime Aerial Tracking (CVPR2022) Junjie Ye, Changhong Fu, Guangze Zheng, Danda Pani Paudel, and Guang Chen. Uns

Intelligent Vision for Robotics in Complex Environment 91 Dec 30, 2022
[NeurIPS 2021] Official implementation of paper "Learning to Simulate Self-driven Particles System with Coordinated Policy Optimization".

Code for Coordinated Policy Optimization Webpage | Code | Paper | Talk (English) | Talk (Chinese) Hi there! This is the source code of the paper “Lear

DeciForce: Crossroads of Machine Perception and Autonomy 81 Dec 19, 2022
Monitor your ML jobs on mobile devices📱, especially for Google Colab / Kaggle

TF Watcher TF Watcher is a simple to use Python package and web app which allows you to monitor 👀 your Machine Learning training or testing process o

Rishit Dagli 54 Nov 01, 2022
Implementation of paper "DeepTag: A General Framework for Fiducial Marker Design and Detection"

Implementation of paper DeepTag: A General Framework for Fiducial Marker Design and Detection. Project page: https://herohuyongtao.github.io/research/

Yongtao Hu 46 Dec 12, 2022
Template repository to build PyTorch projects from source on any version of PyTorch/CUDA/cuDNN.

The Ultimate PyTorch Source-Build Template Translations: 한국어 TL;DR PyTorch built from source can be x4 faster than a naïve PyTorch install. This repos

Joonhyung Lee/이준형 651 Dec 12, 2022
Record radiologists' eye gaze when they are labeling images.

Record radiologists' eye gaze when they are labeling images. Read for installation, usage, and deep learning examples. Why use MicEye Versatile As a l

24 Nov 03, 2022
Which Style Makes Me Attractive? Interpretable Control Discovery and Counterfactual Explanation on StyleGAN

Interpretable Control Exploration and Counterfactual Explanation (ICE) on StyleGAN Which Style Makes Me Attractive? Interpretable Control Discovery an

Bo Li 11 Dec 01, 2022
Implementation of Bagging and AdaBoost Algorithm

Bagging-and-AdaBoost Implementation of Bagging and AdaBoost Algorithm Dataset Red Wine Quality Data Sets For simplicity, we will have 2 classes of win

Zechen Ma 1 Nov 01, 2021
Architecture Patterns with Python (TDD, DDD, EDM)

architecture-traning Architecture Patterns with Python (TDD, DDD, EDM) Chapter 5. 높은 기어비와 낮은 기어비의 TDD 5.2 도메인 계층 테스트를 서비스 계층으로 옮겨야 하는가? 도메인 계층 테스트 def

minsung sim 2 Mar 04, 2022
Contrastive Language-Image Pretraining

CLIP [Blog] [Paper] [Model Card] [Colab] CLIP (Contrastive Language-Image Pre-Training) is a neural network trained on a variety of (image, text) pair

OpenAI 11.5k Jan 08, 2023
An implementation of "MixHop: Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing" (ICML 2019).

MixHop and N-GCN ⠀ A PyTorch implementation of "MixHop: Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing" (ICML 2019)

Benedek Rozemberczki 393 Dec 13, 2022