基于PaddlePaddle实现的声纹识别系统
前言
本项目使用了EcapaTdnn、ResNetSE、ERes2Net、CAM++等多种先进的声纹识别模型，不排除以后会支持更多模型，同时本项目也支持了MelSpectrogram、Spectrogram、MFCC、Fbank等多种数据预处理方法，使用了ArcFace Loss，ArcFace loss：Additive Angular Margin Loss（加性角度间隔损失函数），对应项目中的AAMLoss，对特征向量和权重归一化，对θ加上角度间隔m，角度间隔比余弦间隔在对角度的影响更加直接，除此之外，还支持AMLoss、ARMLoss、CELoss等多种损失函数。
源码地址：VoiceprintRecognition-PaddlePaddle
使用环境：
- Anaconda 3
- Python 3.8
- PaddlePaddle 2.4.1
- Windows 10 or Ubuntu 18.04
项目特性
- 支持模型：EcapaTdnn、TDNN、Res2Net、ResNetSE、ERes2Net、CAM++
- 支持池化层：AttentiveStatsPool(ASP)、SelfAttentivePooling(SAP)、TemporalStatisticsPooling(TSP)、TemporalAveragePooling(TAP)、TemporalStatsPool(TSTP)
- 支持损失函数：AAMLoss、AMLoss、ARMLoss、CELoss
- 支持预处理方法：MelSpectrogram、Spectrogram、MFCC、Fbank
模型论文：
- EcapaTdnn：ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker Verification
- TDNN：Prediction of speech intelligibility with DNN-based performance measures
- Res2Net：Res2Net: A New Multi-scale Backbone Architecture
- ResNetSE：Squeeze-and-Excitation Networks
- CAMPPlus：CAM++: A Fast and Efficient Network for Speaker Verification Using Context-Aware Masking
- ERes2Net：An Enhanced Res2Net with Local and Global Feature Fusion for Speaker Verification
模型下载
| 模型 | Params(M) | 预处理方法 | 数据集 | train speakers | threshold | EER | MinDCF |
|---|---|---|---|---|---|---|---|
| CAM++ | 7.5 | Fbank | CN-Celeb | 2796 | 0.25 | 0.09485 | 0.56214 |
| ERes2Net | 8.2 | Fbank | CN-Celeb | 2796 | 0.22 | 0.09637 | 0.52627 |
| ResNetSE | 10.7 | Fbank | CN-Celeb | 2796 | |||
| EcapaTdnn | 6.7 | Fbank | CN-Celeb | 2796 | 0.25 | 0.10465 | 0.58521 |
| TDNN | 3.2 | Fbank | CN-Celeb | 2796 | 0.23 | 0.11804 | 0.61070 |
| Res2Net | 7.2 | Fbank | CN-Celeb | 2796 | 0.18 | 0.14126 | 0.68511 |
| ERes2Net | 8.2 | Fbank | 其他数据集 | 20W | 0.36 | 0.02939 | 0.18355 |
| CAM++ | 7.5 | Flank | 其他数据集 | 20W | 0.29 | 0.04768 | 0.31429 |
说明：
- 评估的测试集为CN-Celeb的测试集，包含196个说话人。
安装环境
- 首先安装的是PaddlePaddle的GPU版本，如果已经安装过了，请跳过。
conda install paddlepaddle-gpu==2.4.1 cudatoolkit=10.2 --channel https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/Paddle/
- 安装ppvector库。
使用pip安装，命令如下：
python -m pip install ppvector -U -i https://pypi.tuna.tsinghua.edu.cn/simple
建议源码安装，源码安装能保证使用最新代码。
git clone https://github.com/yeyupiaoling/VoiceprintRecognition_PaddlePaddle.git
cd VoiceprintRecognition_PaddlePaddle/
python setup.py install
修改预处理方法
配置文件中默认使用的是Fbank预处理方法，如果要使用其他预处理方法，可以修改配置文件中的安装下面方式修改，具体的值可以根据自己情况修改。如果不清楚如何设置参数，可以直接删除该部分，直接使用默认值。
# 数据预处理参数
preprocess_conf:
# 音频预处理方法，支持：LogMelSpectrogram、MelSpectrogram、Spectrogram、MFCC、Fbank
feature_method: 'Fbank'
# 设置API参数，更参数查看对应API，不清楚的可以直接删除该部分，直接使用默认值
method_args:
sr: 16000
n_mels: 80
训练模型
使用train.py训练模型，本项目支持多个音频预处理方式，通过configs/ecapa_tdnn.yml配置文件的参数preprocess_conf.feature_method可以指定，MelSpectrogram为梅尔频谱，Spectrogram为语谱图，MFCC梅尔频谱倒谱系数。通过参数augment_conf_path可以指定数据增强方式。训练过程中，会使用VisualDL保存训练日志，通过启动VisualDL可以随时查看训练结果，启动命令visualdl --logdir=log --host 0.0.0.0
# 单卡训练
CUDA_VISIBLE_DEVICES=0 python train.py
# 多卡训练
python -m paddle.distributed.launch --gpus '0,1' train.py
训练输出日志：
[2023-08-05 09:52:06.497988 INFO ] utils:print_arguments:13 - ----------- 额外配置参数 -----------
[2023-08-05 09:52:06.498094 INFO ] utils:print_arguments:15 - configs: configs/ecapa_tdnn.yml
[2023-08-05 09:52:06.498149 INFO ] utils:print_arguments:15 - do_eval: True
[2023-08-05 09:52:06.498191 INFO ] utils:print_arguments:15 - local_rank: 0
[2023-08-05 09:52:06.498230 INFO ] utils:print_arguments:15 - pretrained_model: None
[2023-08-05 09:52:06.498269 INFO ] utils:print_arguments:15 - resume_model: None
[2023-08-05 09:52:06.498306 INFO ] utils:print_arguments:15 - save_model_path: models/
[2023-08-05 09:52:06.498342 INFO ] utils:print_arguments:15 - use_gpu: True
[2023-08-05 09:52:06.498378 INFO ] utils:print_arguments:16 - ------------------------------------------------
[2023-08-05 09:52:06.513761 INFO ] utils:print_arguments:18 - ----------- 配置文件参数 -----------
[2023-08-05 09:52:06.513906 INFO ] utils:print_arguments:21 - dataset_conf:
[2023-08-05 09:52:06.513957 INFO ] utils:print_arguments:24 - dataLoader:
[2023-08-05 09:52:06.513995 INFO ] utils:print_arguments:26 - batch_size: 64
[2023-08-05 09:52:06.514031 INFO ] utils:print_arguments:26 - num_workers: 4
[2023-08-05 09:52:06.514066 INFO ] utils:print_arguments:28 - do_vad: False
[2023-08-05 09:52:06.514101 INFO ] utils:print_arguments:28 - enroll_list: dataset/enroll_list.txt
[2023-08-05 09:52:06.514135 INFO ] utils:print_arguments:24 - eval_conf:
[2023-08-05 09:52:06.514169 INFO ] utils:print_arguments:26 - batch_size: 1
[2023-08-05 09:52:06.514203 INFO ] utils:print_arguments:26 - max_duration: 20
[2023-08-05 09:52:06.514237 INFO ] utils:print_arguments:28 - max_duration: 3
[2023-08-05 09:52:06.514274 INFO ] utils:print_arguments:28 - min_duration: 0.5
[2023-08-05 09:52:06.514308 INFO ] utils:print_arguments:28 - noise_aug_prob: 0.2
[2023-08-05 09:52:06.514342 INFO ] utils:print_arguments:28 - noise_dir: dataset/noise
[2023-08-05 09:52:06.514374 INFO ] utils:print_arguments:28 - num_speakers: 3242
[2023-08-05 09:52:06.514408 INFO ] utils:print_arguments:28 - sample_rate: 16000
[2023-08-05 09:52:06.514441 INFO ] utils:print_arguments:28 - speed_perturb: True
[2023-08-05 09:52:06.514475 INFO ] utils:print_arguments:28 - target_dB: -20
[2023-08-05 09:52:06.514508 INFO ] utils:print_arguments:28 - train_list: dataset/train_list.txt
[2023-08-05 09:52:06.514542 INFO ] utils:print_arguments:28 - trials_list: dataset/trials_list.txt
[2023-08-05 09:52:06.514575 INFO ] utils:print_arguments:28 - use_dB_normalization: True
[2023-08-05 09:52:06.514609 INFO ] utils:print_arguments:21 - loss_conf:
[2023-08-05 09:52:06.514643 INFO ] utils:print_arguments:24 - args:
[2023-08-05 09:52:06.514678 INFO ] utils:print_arguments:26 - easy_margin: False
[2023-08-05 09:52:06.514713 INFO ] utils:print_arguments:26 - margin: 0.2
[2023-08-05 09:52:06.514746 INFO ] utils:print_arguments:26 - scale: 32
[2023-08-05 09:52:06.514779 INFO ] utils:print_arguments:24 - margin_scheduler_args:
[2023-08-05 09:52:06.514814 INFO ] utils:print_arguments:26 - final_margin: 0.3
[2023-08-05 09:52:06.514848 INFO ] utils:print_arguments:28 - use_loss: AAMLoss
[2023-08-05 09:52:06.514882 INFO ] utils:print_arguments:28 - use_margin_scheduler: True
[2023-08-05 09:52:06.514915 INFO ] utils:print_arguments:21 - model_conf:
[2023-08-05 09:52:06.514950 INFO ] utils:print_arguments:24 - backbone:
[2023-08-05 09:52:06.514984 INFO ] utils:print_arguments:26 - embd_dim: 192
[2023-08-05 09:52:06.515017 INFO ] utils:print_arguments:26 - pooling_type: ASP
[2023-08-05 09:52:06.515050 INFO ] utils:print_arguments:24 - classifier:
[2023-08-05 09:52:06.515084 INFO ] utils:print_arguments:26 - num_blocks: 0
[2023-08-05 09:52:06.515118 INFO ] utils:print_arguments:21 - optimizer_conf:
[2023-08-05 09:52:06.515154 INFO ] utils:print_arguments:28 - learning_rate: 0.001
[2023-08-05 09:52:06.515188 INFO ] utils:print_arguments:28 - optimizer: Adam
[2023-08-05 09:52:06.515221 INFO ] utils:print_arguments:28 - scheduler: CosineAnnealingLR
[2023-08-05 09:52:06.515254 INFO ] utils:print_arguments:28 - scheduler_args: None
[2023-08-05 09:52:06.515289 INFO ] utils:print_arguments:28 - weight_decay: 1e-06
[2023-08-05 09:52:06.515323 INFO ] utils:print_arguments:21 - preprocess_conf:
[2023-08-05 09:52:06.515357 INFO ] utils:print_arguments:28 - feature_method: MelSpectrogram
[2023-08-05 09:52:06.515390 INFO ] utils:print_arguments:24 - method_args:
[2023-08-05 09:52:06.515426 INFO ] utils:print_arguments:26 - f_max: 14000.0
[2023-08-05 09:52:06.515460 INFO ] utils:print_arguments:26 - f_min: 50.0
[2023-08-05 09:52:06.515493 INFO ] utils:print_arguments:26 - hop_length: 320
[2023-08-05 09:52:06.515527 INFO ] utils:print_arguments:26 - n_fft: 1024
[2023-08-05 09:52:06.515560 INFO ] utils:print_arguments:26 - n_mels: 64
[2023-08-05 09:52:06.515593 INFO ] utils:print_arguments:26 - sample_rate: 16000
[2023-08-05 09:52:06.515626 INFO ] utils:print_arguments:26 - win_length: 1024
[2023-08-05 09:52:06.515660 INFO ] utils:print_arguments:21 - train_conf:
[2023-08-05 09:52:06.515694 INFO ] utils:print_arguments:28 - log_interval: 100
[2023-08-05 09:52:06.515728 INFO ] utils:print_arguments:28 - max_epoch: 30
[2023-08-05 09:52:06.515761 INFO ] utils:print_arguments:30 - use_model: EcapaTdnn
[2023-08-05 09:52:06.515794 INFO ] utils:print_arguments:31 - ------------------------------------------------
----------------------------------------------------------------------------------------
Layer (type) Input Shape Output Shape Param #
========================================================================================
Conv1D-2 [[1, 64, 102]] [1, 512, 98] 164,352
Conv1d-1 [[1, 64, 98]] [1, 512, 98] 0
ReLU-1 [[1, 512, 98]] [1, 512, 98] 0
BatchNorm1D-2 [[1, 512, 98]] [1, 512, 98] 2,048
BatchNorm1d-1 [[1, 512, 98]] [1, 512, 98] 0
TDNNBlock-1 [[1, 64, 98]] [1, 512, 98] 0
Conv1D-4 [[1, 512, 98]] [1, 512, 98] 262,656
Conv1d-3 [[1, 512, 98]] [1, 512, 98] 0
ReLU-2 [[1, 512, 98]] [1, 512, 98] 0
BatchNorm1D-4 [[1, 512, 98]] [1, 512, 98] 2,048
BatchNorm1d-3 [[1, 512, 98]] [1, 512, 98] 0
TDNNBlock-2 [[1, 512, 98]] [1, 512, 98] 0
··········································
SEBlock-3 [[1, 512, 98], None] [1, 512, 98] 0
SERes2NetBlock-3 [[1, 512, 98]] [1, 512, 98] 0
Conv1D-70 [[1, 1536, 98]] [1, 1536, 98] 2,360,832
Conv1d-69 [[1, 1536, 98]] [1, 1536, 98] 0
ReLU-32 [[1, 1536, 98]] [1, 1536, 98] 0
BatchNorm1D-58 [[1, 1536, 98]] [1, 1536, 98] 6,144
BatchNorm1d-57 [[1, 1536, 98]] [1, 1536, 98] 0
TDNNBlock-29 [[1, 1536, 98]] [1, 1536, 98] 0
Conv1D-72 [[1, 4608, 98]] [1, 128, 98] 589,952
Conv1d-71 [[1, 4608, 98]] [1, 128, 98] 0
ReLU-33 [[1, 128, 98]] [1, 128, 98] 0
BatchNorm1D-60 [[1, 128, 98]] [1, 128, 98] 512
BatchNorm1d-59 [[1, 128, 98]] [1, 128, 98] 0
TDNNBlock-30 [[1, 4608, 98]] [1, 128, 98] 0
Tanh-1 [[1, 128, 98]] [1, 128, 98] 0
Conv1D-74 [[1, 128, 98]] [1, 1536, 98] 198,144
Conv1d-73 [[1, 128, 98]] [1, 1536, 98] 0
AttentiveStatisticsPooling-1 [[1, 1536, 98]] [1, 3072, 1] 0
BatchNorm1D-62 [[1, 3072, 1]] [1, 3072, 1] 12,288
BatchNorm1d-61 [[1, 3072, 1]] [1, 3072, 1] 0
Conv1D-76 [[1, 3072, 1]] [1, 192, 1] 590,016
Conv1d-75 [[1, 3072, 1]] [1, 192, 1] 0
EcapaTdnn-1 [[1, 98, 64]] [1, 192] 0
SpeakerIdentification-1 [[1, 192]] [1, 9726] 1,867,392
========================================================================================
Total params: 8,039,808
Trainable params: 8,020,480
Non-trainable params: 19,328
----------------------------------------------------------------------------------------
Input size (MB): 0.02
Forward/backward pass size (MB): 35.60
Params size (MB): 30.67
Estimated Total Size (MB): 66.30
----------------------------------------------------------------------------------------
[2023-08-05 09:52:08.084231 INFO ] trainer:train:388 - 训练数据：874175
[2023-08-05 09:52:09.186542 INFO ] trainer:__train_epoch:334 - Train epoch: [1/30], batch: [0/13659], loss: 11.95824, accuracy: 0.00000, learning rate: 0.00100000, speed: 58.09 data/sec, eta: 5 days, 5:24:08
[2023-08-05 09:52:22.477905 INFO ] trainer:__train_epoch:334 - Train epoch: [1/30], batch: [100/13659], loss: 10.35675, accuracy: 0.00278, learning rate: 0.00100000, speed: 481.65 data/sec, eta: 15:07:15
[2023-08-05 09:52:35.948581 INFO ] trainer:__train_epoch:334 - Train epoch: [1/30], batch: [200/13659], loss: 10.22089, accuracy: 0.00505, learning rate: 0.00100000, speed: 475.27 data/sec, eta: 15:19:12
[2023-08-05 09:52:49.249098 INFO ] trainer:__train_epoch:334 - Train epoch: [1/30], batch: [300/13659], loss: 10.00268, accuracy: 0.00706, learning rate: 0.00100000, speed: 481.45 data/sec, eta: 15:07:11
[2023-08-05 09:53:03.716015 INFO ] trainer:__train_epoch:334 - Train epoch: [1/30], batch: [400/13659], loss: 9.76052, accuracy: 0.00830, learning rate: 0.00100000, speed: 442.74 data/sec, eta: 16:26:16
[2023-08-05 09:53:18.258807 INFO ] trainer:__train_epoch:334 - Train epoch: [1/30], batch: [500/13659], loss: 9.50189, accuracy: 0.01060, learning rate: 0.00100000, speed: 440.46 data/sec, eta: 16:31:08
[2023-08-05 09:53:31.618354 INFO ] trainer:__train_epoch:334 - Train epoch: [1/30], batch: [600/13659], loss: 9.26083, accuracy: 0.01256, learning rate: 0.00100000, speed: 479.50 data/sec, eta: 15:10:12
[2023-08-05 09:53:45.439642 INFO ] trainer:__train_epoch:334 - Train epoch: [1/30], batch: [700/13659], loss: 9.03548, accuracy: 0.01449, learning rate: 0.00099999, speed: 463.63 data/sec, eta: 15:41:08
VisualDL页面：
评估模型
训练结束之后会保存预测模型，我们用预测模型来预测测试集中的音频特征，然后使用音频特征进行两两对比，计算EER和MinDCF。
python eval.py
输出类似如下：
······
------------------------------------------------
W0425 08:27:32.057426 17654 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.5, Driver API Version: 11.6, Runtime API Version: 10.2
W0425 08:27:32.065165 17654 device_context.cc:465] device: 0, cuDNN Version: 7.6.
[2023-03-16 20:20:47.195908 INFO ] trainer:evaluate:341 - 成功加载模型：models/EcapaTdnn_Fbank/best_model/model.pth
100%|███████████████████████████| 84/84 [00:28<00:00, 2.95it/s]
开始两两对比音频特征...
100%|███████████████████████████| 5332/5332 [00:05<00:00, 1027.83it/s]
评估消耗时间：65s，threshold：0.26，EER: 0.14739, MinDCF: 0.41999
声纹对比
下面开始实现声纹对比，创建infer_contrast.py程序，编写infer()函数，在编写模型的时候，模型是有两个输出的，第一个是模型的分类输出，第二个是音频特征输出。所以在这里要输出的是音频的特征值，有了音频的特征值就可以做声纹识别了。我们输入两个语音，通过预测函数获取他们的特征数据，使用这个特征数据可以求他们的对角余弦值，得到的结果可以作为他们相识度。对于这个相识度的阈值threshold，读者可以根据自己项目的准确度要求进行修改。
python infer_contrast.py --audio_path1=audio/a_1.wav --audio_path2=audio/b_2.wav
输出类似如下：
[2023-04-02 18:30:48.009149 INFO ] utils:print_arguments:13 - ----------- 额外配置参数 -----------
[2023-04-02 18:30:48.009149 INFO ] utils:print_arguments:15 - audio_path1: dataset/a_1.wav
[2023-04-02 18:30:48.009149 INFO ] utils:print_arguments:15 - audio_path2: dataset/b_2.wav
[2023-04-02 18:30:48.009149 INFO ] utils:print_arguments:15 - configs: configs/ecapa_tdnn.yml
[2023-04-02 18:30:48.009149 INFO ] utils:print_arguments:15 - model_path: models/EcapaTdnn_Fbank/best_model/
[2023-04-02 18:30:48.009149 INFO ] utils:print_arguments:15 - threshold: 0.6
[2023-04-02 18:30:48.009149 INFO ] utils:print_arguments:15 - use_gpu: True
[2023-04-02 18:30:48.009149 INFO ] utils:print_arguments:16 - ------------------------------------------------
······································································
W0425 08:29:10.006249 21121 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.5, Driver API Version: 11.6, Runtime API Version: 10.2
W0425 08:29:10.008555 21121 device_context.cc:465] device: 0, cuDNN Version: 7.6.
成功加载模型参数和优化方法参数：models/ecapa_tdnn/model.pdparams
audio/a_1.wav 和 audio/b_2.wav 不是同一个人，相似度为：-0.09565544128417969
声纹识别
在上面的声纹对比的基础上，我们创建infer_recognition.py实现声纹识别。同样是使用上面声纹对比的infer()预测函数，通过这两个同样获取语音的特征数据。 不同的是笔者增加了load_audio_db()和register()，以及recognition()，第一个函数是加载声纹库中的语音数据，这些音频就是相当于已经注册的用户，他们注册的语音数据会存放在这里，如果有用户需要通过声纹登录，就需要拿到用户的语音和语音库中的语音进行声纹对比，如果对比成功，那就相当于登录成功并且获取用户注册时的信息数据。第二个函数register()其实就是把录音保存在声纹库中，同时获取该音频的特征添加到待对比的数据特征中。最后recognition()函数中，这个函数就是将输入的语音和语音库中的语音一一对比。
有了上面的声纹识别的函数，读者可以根据自己项目的需求完成声纹识别的方式，例如笔者下面提供的是通过录音来完成声纹识别。首先必须要加载语音库中的语音，语音库文件夹为audio_db，然后用户回车后录音3秒钟，然后程序会自动录音，并使用录音到的音频进行声纹识别，去匹配语音库中的语音，获取用户的信息。通过这样方式，读者也可以修改成通过服务请求的方式完成声纹识别，例如提供一个API供APP调用，用户在APP上通过声纹登录时，把录音到的语音发送到后端完成声纹识别，再把结果返回给APP，前提是用户已经使用语音注册，并成功把语音数据存放在audio_db文件夹中。
python infer_recognition.py
输出类似如下：
[2023-04-02 18:31:20.521040 INFO ] utils:print_arguments:13 - ----------- 额外配置参数 -----------
[2023-04-02 18:31:20.521040 INFO ] utils:print_arguments:15 - audio_db_path: audio_db/
[2023-04-02 18:31:20.521040 INFO ] utils:print_arguments:15 - configs: configs/ecapa_tdnn.yml
[2023-04-02 18:31:20.521040 INFO ] utils:print_arguments:15 - model_path: models/EcapaTdnn_Fbank/best_model/
[2023-04-02 18:31:20.521040 INFO ] utils:print_arguments:15 - record_seconds: 3
[2023-04-02 18:31:20.521040 INFO ] utils:print_arguments:15 - threshold: 0.6
[2023-04-02 18:31:20.521040 INFO ] utils:print_arguments:15 - use_gpu: True
[2023-04-02 18:31:20.521040 INFO ] utils:print_arguments:16 - ------------------------------------------------
······································································
W0425 08:30:13.257884 23889 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.5, Driver API Version: 11.6, Runtime API Version: 10.2
W0425 08:30:13.260191 23889 device_context.cc:465] device: 0, cuDNN Version: 7.6.
成功加载模型参数和优化方法参数：models/ecapa_tdnn/model.pdparams
Loaded 沙瑞金 audio.
Loaded 李达康 audio.
请选择功能，0为注册音频到声纹库，1为执行声纹识别：0
按下回车键开机录音，录音3秒中：
开始录音......
录音已结束!
请输入该音频用户的名称：夜雨飘零
请选择功能，0为注册音频到声纹库，1为执行声纹识别：1
按下回车键开机录音，录音3秒中：
开始录音......
录音已结束!
识别说话的为：夜雨飘零，相似度为：0.920434
其他版本
- Tensorflow：VoiceprintRecognition-Tensorflow
- Pytorch：VoiceprintRecognition-Pytorch
- Keras：VoiceprintRecognition-Keras
参考资料
- https://github.com/PaddlePaddle/PaddleSpeech
- https://github.com/yeyupiaoling/PaddlePaddle-MobileFaceNets
- https://github.com/yeyupiaoling/PPASR