File size: 4,290 Bytes
bf864fd db722c0 bf864fd 960ea7a 3259f6e edfae8a 3259f6e 391404d 3259f6e 391404d 3259f6e 391404d 3259f6e 391404d 3259f6e bf864fd 960ea7a 3259f6e 960ea7a bf864fd 16c4d0d 391404d 16c4d0d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 |
---
language:
- en
license: mit
tags:
- embeddings
- Speaker
- Verification
- Identification
- NAS
- TDNN
- pytorch
datasets:
- voxceleb1
- voxceleb2
metrics:
- EER
- minDCF:
- p_target: 0.01
---
# 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
```python
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](https://arxiv.org/abs/2103.13581).
## **Citing EfficientTDNN**
Please, cite EfficientTDNN if you use it for your research or business.
```bibtex
@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}}
```
|