“Torch speaker”版本间的差异
(→工具代码和脚本) |
(→训练 trick 相关) |
||
(相同用户的20个中间修订版本未显示) | |||
第1行: | 第1行: | ||
− | PyTorch Speaker 一个基于 PyTorch | + | PyTorch Speaker 一个基于 PyTorch 编写的说话人识别'''不止限用于'''科研的工具包。 |
+ | |||
== PyTorch Speaker 简介 == | == PyTorch Speaker 简介 == | ||
− | '''PyTorch Speaker''' | + | '''PyTorch Speaker''' 是一个由 [https://zyzisyz.github.io/about/ zhangy20]编写基于 PyTorch 编写的说话人识别工具包。 |
+ | |||
+ | '''项目特点''' | ||
+ | |||
+ | * 不依赖于Kaldi,没有使用高级的shell语法 | ||
+ | * 支持离线手机端/嵌入式设备的快速部署 | ||
+ | * 丰富的数据可视化支持 | ||
+ | * 可用于计算机安全领域(Anti-Spoofing和Adversarial Attack)语音相关的实验和科研 | ||
+ | * 对于声学特征的提取是Online的 | ||
+ | |||
+ | '''项目可能存在的缺陷''' | ||
+ | |||
+ | * 模型的训练对硬盘的IO性能要求比较高(为了在训练时达到最高的显卡利用率,强烈推荐使用固态硬盘存放数据) | ||
说话人识别(Speaker Recognition, SRE)技术,也称为声纹识别(Voiceprint Recognition, VPR)技术属于生物特征识别技术的一种,是一项根据语音信号中反映说话人生理和行为特征的语音参数(“声纹”),自动识别说话人身份的技术。说话人识别本质上是一类模式识别问题。说话人识别按照具体场景和需求的不同可以分为如下图所示的3个子任务: | 说话人识别(Speaker Recognition, SRE)技术,也称为声纹识别(Voiceprint Recognition, VPR)技术属于生物特征识别技术的一种,是一项根据语音信号中反映说话人生理和行为特征的语音参数(“声纹”),自动识别说话人身份的技术。说话人识别本质上是一类模式识别问题。说话人识别按照具体场景和需求的不同可以分为如下图所示的3个子任务: | ||
第9行: | 第22行: | ||
[[文件:speaker_task.png|700px|center]] | [[文件:speaker_task.png|700px|center]] | ||
+ | <center> | ||
{| border="1" cellspacing="0" cellpadding="5" | {| border="1" cellspacing="0" cellpadding="5" | ||
| 任务中文名称 | | 任务中文名称 | ||
第17行: | 第31行: | ||
| 说话人辨认 | | 说话人辨认 | ||
| Speaker Identification | | Speaker Identification | ||
− | | | + | | 判断某段语音是由若干人中的哪一个人所说,是“1 vs N”的判别 |
− | | Speaker Identification that identify the true speaker from a set of candidates | + | | Speaker Identification that identify the true speaker from a set of candidates. |
|- | |- | ||
| 说话人确认 | | 说话人确认 | ||
| Speaker Verification | | Speaker Verification | ||
− | | | + | | 判断某段语音是否是由指定的某个人所说,是“1 vs 1”的判别 |
| Speaker Verification that tests if an alleged speaker is the true speaker. | | Speaker Verification that tests if an alleged speaker is the true speaker. | ||
|- | |- | ||
第31行: | 第45行: | ||
|- | |- | ||
|} | |} | ||
+ | </center> | ||
− | === | + | === Demo 演示 === |
− | + | === 系统性能 === | |
− | + | ||
− | + | ||
− | === | + | <center> |
+ | {| class=“wikitable” border="1" cellspacing="0" cellpadding="5" style="text-align:center | ||
+ | |- | ||
+ | ! 网络结构 | ||
+ | ! 网络参数量 | ||
+ | ! 损失函数 | ||
+ | ! 数据增强 | ||
+ | ! 训练数据集 | ||
+ | ! 测试数据集 | ||
+ | ! cosine EER | ||
+ | ! cosine minDCF(10-2) | ||
+ | ! cosine minDCF(10-3) | ||
+ | ! 清华网盘链接 | ||
+ | |- | ||
+ | |- | ||
+ | | rowspan="3" | resnet34_TSP | ||
− | == | + | [1, 32, 64, 128, 256] |
+ | | rowspan="3" | 6.4M | ||
+ | | rowspan="3" | amsoftmax | ||
− | + | (margin=0.2, scale=30) | |
− | | | + | | rowspan="3" | NO |
− | | | + | | rowspan="3" | voxceleb1&2-dev |
− | | | + | |
− | | | + | (7205) |
− | | | + | | vox1_clean |
− | | | + | | 1.59% |
− | | | + | | 0.188 |
− | | | + | | 0.334 |
− | | | + | | rowspan="3" | [https://cloud.tsinghua.edu.cn/d/1f4cb05ba2ef4134882f/ download link] |
− | | | + | |- |
+ | | vox1_E_clean | ||
+ | | 0.77% | ||
+ | | 0.099 | ||
+ | | 0.187 | ||
+ | |- | ||
+ | | vox1_H_clean | ||
+ | | 1.40% | ||
+ | | 0.153 | ||
+ | | 0.280 | ||
|- | |- | ||
− | | | + | | rowspan="3" | resnet34_ASP |
− | | | + | |
− | | | + | [1, 32, 64, 128, 256] |
− | | | + | | rowspan="3" | 10.6M |
− | | | + | | rowspan="3" | amsoftmax |
− | | | + | |
− | | | + | (margin=0.2, scale=30) |
− | | | + | | rowspan="3" | NO |
− | | | + | | rowspan="3" | voxceleb1-dev |
− | | | + | |
+ | (1211) | ||
+ | | vox1_clean | ||
+ | | 5.42% | ||
+ | | 0.471 | ||
+ | | 0.550 | ||
+ | | rowspan="3" | [https://cloud.tsinghua.edu.cn/d/cf4bcf48687f4b44bd83/ download link] | ||
+ | |- | ||
+ | | vox1_E_clean | ||
+ | | 0.86% | ||
+ | | 0.100 | ||
+ | | 0.198 | ||
+ | |- | ||
+ | | vox1_H_clean | ||
+ | | 1.46% | ||
+ | | 0.156 | ||
+ | | 0.295 | ||
+ | |- | ||
+ | | rowspan="3" | resnet34_TSP | ||
+ | |||
+ | [1, 32, 64, 128, 256] | ||
+ | | rowspan="3" | 6.4M | ||
+ | | rowspan="3" | amsoftmax | ||
+ | |||
+ | (margin=0.2, scale=30) | ||
+ | | rowspan="3" | NO | ||
+ | | rowspan="3" | voxceleb1-dev | ||
+ | |||
+ | (1211) | ||
+ | | vox1_clean | ||
+ | | 4.37% | ||
+ | | 0.453 | ||
+ | | 0.577 | ||
+ | | rowspan="3" | [https://cloud.tsinghua.edu.cn/d/8e692cfa57b54467a83d/ download link] | ||
+ | |- | ||
+ | | vox1_E_clean | ||
+ | | 0.613% | ||
+ | | 0.063 | ||
+ | | 0.138 | ||
+ | |- | ||
+ | | vox1_H_clean | ||
+ | | 1.03% | ||
+ | | 0.110 | ||
+ | | 0.226 | ||
|} | |} | ||
+ | </center> | ||
=== 项目结构 === | === 项目结构 === | ||
第70行: | 第153行: | ||
├── config/ # 存放yaml配置文件 | ├── config/ # 存放yaml配置文件 | ||
├── docs/ # 存放文档 | ├── docs/ # 存放文档 | ||
+ | ├── examples/# 存放示例 | ||
├── README.md | ├── README.md | ||
├── requirements.txt | ├── requirements.txt | ||
− | ├── scripts/ # | + | ├── scripts/ # 存放数据处理、数据可视化脚本 |
├── setup.py | ├── setup.py | ||
− | ├── tools/ # | + | ├── tools/ # 存放训练、推理、量化部署等脚本 |
└── torch_speaker/ # 模型pipline的主体实现 | └── torch_speaker/ # 模型pipline的主体实现 | ||
├── backbone/ | ├── backbone/ | ||
− | ├── | + | ├── audio/ |
├── loss/ | ├── loss/ | ||
├── score/ | ├── score/ | ||
− | ├── | + | ├── module.py |
└── utils/ | └── utils/ | ||
== 快速安装与上手 == | == 快速安装与上手 == | ||
− | === | + | <center> |
− | + | {| class=“wikitable” border="1" cellspacing="0" cellpadding="5" | |
+ | | 步骤 | ||
+ | | 使用方法 | ||
+ | | 注意事项 | ||
+ | |- | ||
+ | | '''1. Installation''' | ||
+ | | | ||
git clone | git clone | ||
cd torch_speaker | cd torch_speaker | ||
pip install -r requirements.txt | pip install -r requirements.txt | ||
python setup.py develop | python setup.py develop | ||
− | + | | | |
− | + | |- | |
− | + | | '''2. Data preparation''' | |
+ | | | ||
采用pandas构建datlist.csv来实现数据的准备。 | 采用pandas构建datlist.csv来实现数据的准备。 | ||
+ | python3 scripts/build_datalist.py \ | ||
+ | --extension wav \ | ||
+ | --dataset_dir data/train \ | ||
+ | --data_list_path data/train.csv | ||
− | + | python3 scripts/format_trials.py \ | |
− | + | --voxceleb1_root your_voxceleb1_path \ | |
− | + | --src_trials_path your_src_trail_path \ | |
+ | --dst_trials_path your_dst_trail_path | ||
+ | | 语音数据默认为单声道wav文件,采用率为16k。 | ||
+ | |- | ||
+ | | '''3. Training''' | ||
+ | | | ||
+ | python3 tools/train.py \ | ||
+ | --config your_yaml_path | ||
+ | | | ||
+ | |- | ||
+ | | '''4. Log Visualization''' | ||
+ | | | ||
+ | tensorboard --log_dir your_log_dir/ --bind_all | ||
+ | | [[文件:eer_log.png|250px|center]] | ||
+ | |- | ||
+ | | '''5. Evaluation''' | ||
+ | | | ||
+ | python3 tools/evaluate.py \ | ||
+ | --config your_yaml_path \ | ||
+ | --trial_path your_trial_path \ | ||
+ | --checkpoint_path your_checkpoint_path | ||
+ | | | ||
+ | |- | ||
+ | | '''6. Export''' | ||
+ | | | ||
+ | | | ||
+ | |} | ||
+ | </center> | ||
== 语音声学特征的提取 == | == 语音声学特征的提取 == | ||
− | + | 由于 torchaudio 库中可能存在潜在的bug(经常各种warning),且未来我计划采用小波分析的方法(wavelet)进行一些其他的实验,因此提取声学体征的代码也是完全采用PyTorch手动实现(没有依赖于其他第三方库)。 | |
+ | |||
+ | [[文件:spec.png|850px|center]] | ||
其中,值得注意的是: | 其中,值得注意的是: | ||
* kaldi所实现的特征提取是offline的,采用PyTorch可以实现online的特征提取; | * kaldi所实现的特征提取是offline的,采用PyTorch可以实现online的特征提取; | ||
− | * Mel-Spectrogram 也叫Fbank,或是FilterBank,在Kaldi中叫Fbank中比较多,在TTS和VC中Mel- | + | * Mel-Spectrogram 也叫Fbank,或是FilterBank,在Kaldi中叫Fbank中比较多,在TTS和VC中Mel-Spectrogram叫的比较多; |
* 由我实现的特征提取代码,kaldi,librosa,torchaudio即便是在相同配置参数的情况下,所提取的到的结果都不同。 | * 由我实现的特征提取代码,kaldi,librosa,torchaudio即便是在相同配置参数的情况下,所提取的到的结果都不同。 | ||
− | + | <center> | |
{| border="1" cellspacing="0" cellpadding="5" | {| border="1" cellspacing="0" cellpadding="5" | ||
| 特征名 | | 特征名 | ||
第139行: | 第263行: | ||
| | | | ||
|} | |} | ||
+ | </center> | ||
== 常用backbone与实现 == | == 常用backbone与实现 == | ||
第145行: | 第270行: | ||
=== ResNet和其变种 === | === ResNet和其变种 === | ||
+ | |||
+ | ==== ResNet ==== | ||
+ | |||
+ | ==== SENet ==== | ||
+ | |||
+ | ==== Res2Net ==== | ||
+ | |||
+ | [[文件:res2net.png|center|400px]] | ||
=== TDNN和其变种 === | === TDNN和其变种 === | ||
+ | |||
+ | ==== TDNN(x-vector) ==== | ||
+ | |||
+ | ==== ECAPA-TDNN ==== | ||
+ | |||
+ | ECAPA-TDNN architecture is based on the popular x-vector topology and it introduces several enhancements to create more robust speaker embeddings. | ||
+ | |||
+ | [[文件:ECAPA TDNN.jpeg|center|300px]] | ||
+ | |||
+ | The pooling layer uses a channel-and context-dependent attention mechanism, which allows the network to attend different frames per channel. 1-dimensional SqueezeExcitation (SE) blocks rescale the channels of the intermediate frame-level feature maps to insert global context information in the locally operating convolutional blocks. Next, the integration of 1-dimensional Res2-blocks improves performance while simultaneously reducing the total parameter count by using grouped convolutions in a hierarchical way. | ||
+ | |||
+ | Finally, Multi-layer Feature Aggregation (MFA) merges complementary information before the statistics pooling by concatenating the final frame-level feature map with an intermediate feature maps of preceding layers. | ||
+ | |||
+ | The network is trained by optimizing the AAM-softmax loss on the speaker identities in the training corpus. The AAM-softmax is a powerful enhancement compared to the regular softmax loss in the context of fine-grained classification and verification problems. It directly optimizes the cosine distance between the speaker embeddings. | ||
+ | |||
+ | The model turned out to work amazingly well for speaker verification and speaker diarization. | ||
+ | |||
+ | === VGG和其变种 === | ||
+ | |||
+ | ==== VGG ==== | ||
+ | |||
+ | ==== RepVGG ==== | ||
+ | |||
+ | == 常用pooling layer == | ||
+ | |||
+ | <center> | ||
+ | {| class=“wikitable” border="1" cellspacing="0" cellpadding="5" | ||
+ | | '''Pooling Layer''' | ||
+ | | | ||
+ | | | ||
+ | |- | ||
+ | | TSP | ||
+ | | | ||
+ | | | ||
+ | |- | ||
+ | | TAP | ||
+ | | | ||
+ | | | ||
+ | |- | ||
+ | | ASP | ||
+ | | | ||
+ | | | ||
+ | |- | ||
+ | | SAP | ||
+ | | | ||
+ | | | ||
+ | |} | ||
+ | </center> | ||
== 常用Loss == | == 常用Loss == | ||
+ | |||
+ | <center> | ||
+ | {| class=“wikitable” border="1" cellspacing="0" cellpadding="5" | ||
+ | | Loss | ||
+ | | 描述 | ||
+ | | 计算公式 | ||
+ | | 代码实现 | ||
+ | |- | ||
+ | | softmax | ||
+ | | | ||
+ | | | ||
+ | | | ||
+ | |- | ||
+ | | Triplet Loss | ||
+ | | | ||
+ | Triplet Loss基本思路是构造一个三元组,由anchor、positive 和 negative 组成,其中 anchor 和 positive 表示来自于同一个人的不同声音,negative 表示来自不同的人的声音,然后,用大量标注好的三元组作为网络输入,训练DNN参数。其优点在于直接使用embeddings之间的相似度作为优化的成本函数,最大化 anchor 和 positive的相似度,同时最小化 anchor和 negative 的相似度。这样,在提取了说话者的 embedding 之后,说话人识别任务就可以简单地通过相似度计算实现。 | ||
+ | | | ||
+ | | | ||
+ | |- | ||
+ | | AM-softmax | ||
+ | | | ||
+ | Kaldi搭建的声纹系统在模型训练中大多使用Softmax损失函数,但 Softmax 损失函数并不能增大类内紧凑性和类间分离性,为了增强embedding的判别性 | ||
+ | | | ||
+ | | | ||
+ | |- | ||
+ | | AAM-softmax | ||
+ | | | ||
+ | | | ||
+ | | | ||
+ | |} | ||
+ | </center> | ||
== 后端打分 == | == 后端打分 == | ||
− | == data loader == | + | <center> |
+ | {| class=“wikitable” border="1" cellspacing="0" cellpadding="5" | ||
+ | | 打分方法 | ||
+ | | 功能描述 | ||
+ | | 计算公式 | ||
+ | | 代码实现 | ||
+ | |- | ||
+ | | cosine | ||
+ | | | ||
+ | | | ||
+ | | | ||
+ | |- | ||
+ | | GPLDA | ||
+ | | | ||
+ | | | ||
+ | | | ||
+ | |- | ||
+ | | LDA-GPLDA | ||
+ | | | ||
+ | | | ||
+ | | | ||
+ | |} | ||
+ | </center> | ||
+ | |||
+ | == 训练 trick 相关 == | ||
+ | |||
+ | <center> | ||
+ | {| class=“wikitable” border="1" cellspacing="0" cellpadding="5" | ||
+ | | tricks | ||
+ | | 注意事项 | ||
+ | |- | ||
+ | | distributed modes | ||
+ | | | ||
+ | * Data Parallel (accelerator=’dp’) (multiple-gpus, 1 machine) | ||
+ | * DistributedDataParallel (accelerator=’ddp’) (multiple-gpus across many machines (python script based)). | ||
+ | * DistributedDataParallel (accelerator=’ddp_spawn’) (multiple-gpus across many machines (spawn based)). | ||
+ | * DistributedDataParallel 2 (accelerator=’ddp2’) (DP in a machine, DDP across machines). | ||
+ | * Horovod (accelerator=’horovod’) (multi-machine, multi-gpu, configured at runtime) | ||
+ | * TPUs (tpu_cores=8|x) (tpu or TPU pod) | ||
+ | |- | ||
+ | | warm-up | ||
+ | | | ||
+ | |- | ||
+ | | learning rate schedule | ||
+ | | | ||
+ | |- | ||
+ | | optimizer | ||
+ | |尽管momentum SDG收敛速度慢于Adam,但是SGD的效果在实验上可能往往好于Adam(这个地方存疑),但在数据量大+做了数据增强后,使用SGD的训练时间实在是太久,综合来看还是使用Adam比较好 | ||
+ | |- | ||
+ | |chunk size | ||
+ | | | ||
+ | chunk size 对模型的训练效果有着非常显著的影响,最好把chunk size改小一点,不要太大。从目前来的实验结果看2秒往往优于2.5秒和3秒,不清楚chunk size更小能不能有更大的提升。 | ||
+ | |- | ||
+ | | 数据增强 | ||
+ | | 数据增强有两种实现思路: | ||
+ | * 每一个batch中的各个数据做随机的去不同增强, | ||
+ | * 每一个batch中的所有数据做相同的数据增强。 | ||
+ | |||
+ | 第一种直觉上可能会比第二种会更好,但实际使用过程中,第一种方法loss收敛速度很慢,我猜测可能在一个batch中做不同的数据增强会导致loss存在一定程度的发散。第二种还在开发中。 | ||
+ | 如果只是想快速验证一些科研想法,推荐不要开启数据增强。目前有文献表明,如果开启数据增强相比于不开启将会花费6倍的训练时间,数据增强应该是在 | ||
+ | * 数据量少,例如只用了voxceleb1 | ||
+ | * 算法和模型验证无误确实有效,需要冲击一个SOTA的模型 | ||
+ | * 服务器可以连续训练好几天不停止只跑这一个模型的情况 | ||
+ | 以上几种情况下使用 | ||
+ | |} | ||
+ | </center> | ||
+ | |||
+ | == data loader 相关 == | ||
=== 数据增强 === | === 数据增强 === | ||
− | === | + | <center> |
+ | {| class=“wikitable” border="1" cellspacing="0" cellpadding="5" | ||
+ | | WavAugment | ||
+ | | | ||
+ | * 加性噪声 | ||
+ | * 音量扰动 | ||
+ | * 加入混响 | ||
+ | * 速度扰动 | ||
+ | |- | ||
+ | | SpecAugment | ||
+ | | | ||
+ | * frequency mask | ||
+ | * time(frame) mask | ||
+ | * frequency swap | ||
+ | * time(frame) swap | ||
+ | |} | ||
+ | </center> | ||
− | == | + | === 训练数据类别均衡 === |
− | + | 在自然情况下,数据往往都会呈现'''长尾分布'''。这种趋势同样出现在从自然科学到社会科学的各个领域各个问题中,直接利用长尾数据来训练的分类和识别系统,往往会对头部数据过拟合,从而在预测时忽略尾部的类别。说话人数据'''类别分布不均衡'''可能会导致训练的模型效果一般,为此需要在dataloader上使用一些预处理的手段实现说话人类别的均衡。 | |
− | === minDCF === | + | 长尾分布的最简单的两类基本方法是: |
+ | * '''重采样(re-sampling)''':对头部类别的欠采样(under-sampling)和对尾部类别的过采样(over-sampling) | ||
+ | * '''重加权(re-weighting)''':重加权则主要体现在分类的loss上 | ||
+ | |||
+ | 这两类方法本质都是利用已知的数据集分布,在学习过程中对数据分布进行反向加权,强化尾部类别的学习,抵消长尾效应。在这个项目中,我采用'''重采样(re-sampling)'''解决数据分布不均衡的问题。 | ||
+ | |||
+ | [[文件:voxceleb_utt.png|1000px|center]] | ||
+ | |||
+ | === 说话人识别常用的数据集 === | ||
+ | |||
+ | == 评价指标相关 == | ||
+ | |||
+ | === 混淆矩阵 === | ||
+ | |||
+ | 对于一个二分类系统,将实例分为正类(Positive)、负类(Negative),则模式分类器有四种分类结果: | ||
+ | |||
+ | * '''TP'''(True Positive):正确的正例,一个实例是正类并且也被判定成正类; | ||
+ | * '''FN'''(False Negative):错误的反例,漏报,本为正类但判定为假类; | ||
+ | * '''FP'''(False Positive):错误的正例,误报,本为假类但判定为正类; | ||
+ | * '''TN'''(True Negative):正确的反例,一个实例是假类并且也被判定成假类。 | ||
+ | |||
+ | === ROC 曲线 === | ||
+ | |||
+ | 对于Speaker Verification任务,评估模型的方法可以是绘制ROC曲线(Receiver Operating Characteristic curve),首先了解以下定义 | ||
+ | |||
+ | * 真阳率(True Positive Rate, '''TPR'''):描述识别出的所有正例占所有正例的比例 | ||
+ | * 假阳率(False Positive Rate, '''FPR'''):描述将负例识别为正例的情况占所有负例的比例 | ||
+ | * 真阴率(True Negative Rate,'''TNR'''):描述识别出的负例占所有负例的比例 | ||
+ | * 其中其中TPR即为敏感度(sensitivity),TNR即为特异度(specificity) | ||
+ | |||
+ | ROC曲线将TPR定义为X轴,将FPR定义为Y轴; | ||
+ | AUC(Area Under Curve)是指ROC曲线下面积,越接近1表示分类器越好。曲线下面积越大,分类的准确性就越高; | ||
+ | 最靠近坐标图左上方的点为灵敏性和特异性均较高的临界值。 | ||
+ | ROC曲线有个很好的特性:'''当测试集中的正负样本的分布变化的时候,ROC曲线能够保持不变。''' | ||
+ | |||
+ | === Equal Error Rate(EER) === | ||
+ | |||
+ | * '''虚警率'''(False alarm)表示负类样本被分为正类样本在所有负类样本中的比例 | ||
+ | * '''漏警率表示'''(Missing alarm)表示正类样本被分为负类样本在所有正类样本中的比例 | ||
+ | |||
+ | === Minimum Detection Cost Function(minDCF) === | ||
+ | |||
+ | === Diarization Error Rate (DER) === | ||
+ | |||
+ | == Anti-Spoofing == | ||
== 对抗样本攻击与防御 == | == 对抗样本攻击与防御 == | ||
+ | |||
+ | 当AI模型/算法设计之初未考虑相关的安全威胁的情况下,AI算法的判断结果容易被恶意攻击者影响,导致AI系统判断失准。其中最主要的安全威胁是闪避攻击,即是指通过修改输人,让AI模型无法对其正确识别。研究表明深度学习系统容易受到精心设计的输人样本的影响,这些输人样本称为对抗样本(Adversarial Examples) | ||
+ | |||
+ | 说话人识别系统,无论是基于深度神经网络的/或是基于传统统计模型的i-vector系统都同样存在这样的问题。 | ||
=== 对抗样本攻击 === | === 对抗样本攻击 === | ||
− | ==== | + | ==== 白盒攻击 ==== |
− | ==== PGD ==== | + | <center> |
+ | {| class=“wikitable” border="1" cellspacing="0" cellpadding="5" | ||
+ | | 攻击方法 | ||
+ | | 计算方法 | ||
+ | | 代码实现 | ||
+ | |- | ||
+ | | BIM | ||
+ | | | ||
+ | | | ||
+ | |- | ||
+ | | PGD | ||
+ | | | ||
+ | | | ||
+ | |} | ||
+ | </center> | ||
+ | |||
+ | ==== 黑盒攻击 ==== | ||
=== 对抗样本防御 === | === 对抗样本防御 === | ||
第176行: | 第535行: | ||
== 工具代码和脚本 == | == 工具代码和脚本 == | ||
+ | === 工具代脚本 === | ||
+ | |||
+ | <center> | ||
{| border="1" cellspacing="0" cellpadding="5" | {| border="1" cellspacing="0" cellpadding="5" | ||
| 脚本名称 | | 脚本名称 | ||
第212行: | 第574行: | ||
|- | |- | ||
| 文档生成 | | 文档生成 | ||
+ | | | ||
+ | | | ||
+ | |- | ||
+ | | format trials | ||
| | | | ||
| | | | ||
|} | |} | ||
+ | </center> | ||
=== 可视化 === | === 可视化 === | ||
− | == | + | <center> |
+ | {| border="1" cellspacing="0" cellpadding="5" | ||
+ | | 功能 | ||
+ | | 效果预览图 | ||
+ | | 代码链接 | ||
+ | |- | ||
+ | | 绘制语谱图(spectrogram) | ||
+ | | [[文件:plot_spec.png|550px|center]] | ||
+ | | | ||
+ | |- | ||
+ | | 绘制3D语谱图(3D-spectrogram) | ||
+ | | [[文件:plot_3d_spec.png|550px|center]] | ||
+ | | | ||
+ | |- | ||
+ | | 绘制ROC曲线 | ||
+ | |[[文件:roc.png|550px|center]] | ||
+ | | | ||
+ | |- | ||
+ | | 绘制PR曲线 | ||
+ | | | ||
+ | | | ||
+ | |- | ||
+ | | 绘制混淆矩阵 | ||
+ | | | ||
+ | | | ||
+ | |} | ||
+ | </center> | ||
− | === | + | === 模型部署 === |
+ | 以 PyTorch 和 TensorFlow 为代表的深度学习框架集成了模型的'''训练'''和'''推理'''两个过程。然而在实际模型的使用中,如果想要在不同类型的平台(云/Edge、CPU/GPU 等)上获得最佳性能,需需要'''调整模型'''(量化、知识蒸馏)和'''使用专门的推理库'''(ONNX,TensorRT)。 | ||
+ | |||
+ | ONNX runtime 是一种用于将 ONNX 模型部署到生产环境的高性能推理引擎。 它针对云和 Edge 进行了优化,适用于 Linux、Windows 和 Mac。 它使用 C++ 编写,还包含 C、Python、C#、Java 和 Javascript (Node.js) API,可在各种环境中使用。 ONNX 运行时同时支持 DNN 和传统 ML 模型,并与不同硬件上的加速器(例如,NVidia GPU 上的 TensorRT、Intel 处理器上的 OpenVINO、Windows 上的 DirectML 等)集成。 通过使用 ONNX 运行时,可以从大量的生产级优化、测试和不断改进中受益。 | ||
+ | |||
+ | === 代码规范 === | ||
+ | |||
+ | == MISC == | ||
+ | |||
+ | <center> | ||
+ | {| border="1" cellspacing="0" cellpadding="5" | ||
+ | | 框架 | ||
+ | | 描述 | ||
+ | |- | ||
+ | | PyTorch | ||
+ | | | ||
PyTorch is a Python package that provides two high-level features: | PyTorch is a Python package that provides two high-level features: | ||
* Tensor computation (like NumPy) with strong GPU acceleration | * Tensor computation (like NumPy) with strong GPU acceleration | ||
* Deep neural networks built on a tape-based autograd system | * Deep neural networks built on a tape-based autograd system | ||
− | + | |- | |
− | + | | PyTorch Lightning | |
− | + | | | |
The goal of PyTorch Lightning is "You do the research. Lightning will do everything else". | The goal of PyTorch Lightning is "You do the research. Lightning will do everything else". | ||
PyTorch Lightning was started by William Falcon while completing his Ph.D. AI research at NYU CILVR and Facebook AI Research, with the vision of making it a foundational part of everyone’s deep learning research code. The framework was designed for professional and academic researchers working in AI, making state of the art AI research techniques, such as TPU training, trivial. | PyTorch Lightning was started by William Falcon while completing his Ph.D. AI research at NYU CILVR and Facebook AI Research, with the vision of making it a foundational part of everyone’s deep learning research code. The framework was designed for professional and academic researchers working in AI, making state of the art AI research techniques, such as TPU training, trivial. | ||
− | + | |- | |
− | + | | ONNX | |
− | + | | | |
Open Neural Network Exchange (ONNX) is an open ecosystem that empowers AI developers to choose the right tools as their project evolves. ONNX provides an open source format for AI models, both deep learning and traditional ML. It defines an extensible computation graph model, as well as definitions of built-in operators and standard data types. | Open Neural Network Exchange (ONNX) is an open ecosystem that empowers AI developers to choose the right tools as their project evolves. ONNX provides an open source format for AI models, both deep learning and traditional ML. It defines an extensible computation graph model, as well as definitions of built-in operators and standard data types. | ||
− | + | |- | |
− | + | | NCNN | |
− | + | | | |
ncnn is a high-performance neural network inference computing framework optimized for mobile platforms. ncnn is deeply considerate about deployment and uses on mobile phones from the beginning of design. ncnn does not have third party dependencies. it is cross-platform, and runs faster than all known open source frameworks on mobile phone cpu. Developers can easily deploy deep learning algorithm models to the mobile platform by using efficient ncnn implementation, create intelligent APPs, and bring the artificial intelligence to your fingertips. ncnn is currently being used in many Tencent applications, such as QQ, Qzone, WeChat, Pitu and so on. | ncnn is a high-performance neural network inference computing framework optimized for mobile platforms. ncnn is deeply considerate about deployment and uses on mobile phones from the beginning of design. ncnn does not have third party dependencies. it is cross-platform, and runs faster than all known open source frameworks on mobile phone cpu. Developers can easily deploy deep learning algorithm models to the mobile platform by using efficient ncnn implementation, create intelligent APPs, and bring the artificial intelligence to your fingertips. ncnn is currently being used in many Tencent applications, such as QQ, Qzone, WeChat, Pitu and so on. | ||
− | + | |- | |
− | + | | YACS | |
− | + | | | |
YACS was created as a lightweight library to define and manage system configurations, such as those commonly found in software designed for scientific experimentation. These "configurations" typically cover concepts like hyperparameters used in training a machine learning model or configurable model hyperparameters, such as the depth of a convolutional neural network. | YACS was created as a lightweight library to define and manage system configurations, such as those commonly found in software designed for scientific experimentation. These "configurations" typically cover concepts like hyperparameters used in training a machine learning model or configurable model hyperparameters, such as the depth of a convolutional neural network. | ||
+ | |- | ||
+ | | Sphinx | ||
+ | | Sphinx is a tool that makes it easy to create intelligent and beautiful documentation, written by Georg Brandl and licensed under the BSD license. | ||
+ | |} | ||
+ | </center> | ||
== 参考 == | == 参考 == | ||
− | |||
− | |||
− | |||
− | |||
− | |||
− |
2021年5月28日 (五) 06:54的最后版本
PyTorch Speaker 一个基于 PyTorch 编写的说话人识别不止限用于科研的工具包。
目录
PyTorch Speaker 简介
PyTorch Speaker 是一个由 zhangy20编写基于 PyTorch 编写的说话人识别工具包。
项目特点
- 不依赖于Kaldi,没有使用高级的shell语法
- 支持离线手机端/嵌入式设备的快速部署
- 丰富的数据可视化支持
- 可用于计算机安全领域(Anti-Spoofing和Adversarial Attack)语音相关的实验和科研
- 对于声学特征的提取是Online的
项目可能存在的缺陷
- 模型的训练对硬盘的IO性能要求比较高(为了在训练时达到最高的显卡利用率,强烈推荐使用固态硬盘存放数据)
说话人识别(Speaker Recognition, SRE)技术,也称为声纹识别(Voiceprint Recognition, VPR)技术属于生物特征识别技术的一种,是一项根据语音信号中反映说话人生理和行为特征的语音参数(“声纹”),自动识别说话人身份的技术。说话人识别本质上是一类模式识别问题。说话人识别按照具体场景和需求的不同可以分为如下图所示的3个子任务:
任务中文名称 | 任务英语名称 | 中文定义 | 英语定义 |
说话人辨认 | Speaker Identification | 判断某段语音是由若干人中的哪一个人所说,是“1 vs N”的判别 | Speaker Identification that identify the true speaker from a set of candidates. |
说话人确认 | Speaker Verification | 判断某段语音是否是由指定的某个人所说,是“1 vs 1”的判别 | Speaker Verification that tests if an alleged speaker is the true speaker. |
说话人追踪 | Speaker Diarization | Speaker Diarization which addresses the problem of “who spoken and when”, which is a process of partitioning a conversation recording into several speech recordings, each of which belongs to a single speaker. |
Demo 演示
系统性能
网络结构 | 网络参数量 | 损失函数 | 数据增强 | 训练数据集 | 测试数据集 | cosine EER | cosine minDCF(10-2) | cosine minDCF(10-3) | 清华网盘链接 |
---|---|---|---|---|---|---|---|---|---|
resnet34_TSP
[1, 32, 64, 128, 256] |
6.4M | amsoftmax
(margin=0.2, scale=30) |
NO | voxceleb1&2-dev
(7205) |
vox1_clean | 1.59% | 0.188 | 0.334 | download link |
vox1_E_clean | 0.77% | 0.099 | 0.187 | ||||||
vox1_H_clean | 1.40% | 0.153 | 0.280 | ||||||
resnet34_ASP
[1, 32, 64, 128, 256] |
10.6M | amsoftmax
(margin=0.2, scale=30) |
NO | voxceleb1-dev
(1211) |
vox1_clean | 5.42% | 0.471 | 0.550 | download link |
vox1_E_clean | 0.86% | 0.100 | 0.198 | ||||||
vox1_H_clean | 1.46% | 0.156 | 0.295 | ||||||
resnet34_TSP
[1, 32, 64, 128, 256] |
6.4M | amsoftmax
(margin=0.2, scale=30) |
NO | voxceleb1-dev
(1211) |
vox1_clean | 4.37% | 0.453 | 0.577 | download link |
vox1_E_clean | 0.613% | 0.063 | 0.138 | ||||||
vox1_H_clean | 1.03% | 0.110 | 0.226 |
项目结构
. ├── config/ # 存放yaml配置文件 ├── docs/ # 存放文档 ├── examples/# 存放示例 ├── README.md ├── requirements.txt ├── scripts/ # 存放数据处理、数据可视化脚本 ├── setup.py ├── tools/ # 存放训练、推理、量化部署等脚本 └── torch_speaker/ # 模型pipline的主体实现 ├── backbone/ ├── audio/ ├── loss/ ├── score/ ├── module.py └── utils/
快速安装与上手
步骤 | 使用方法 | 注意事项 |
1. Installation |
git clone cd torch_speaker pip install -r requirements.txt python setup.py develop |
|
2. Data preparation |
采用pandas构建datlist.csv来实现数据的准备。 python3 scripts/build_datalist.py \ --extension wav \ --dataset_dir data/train \ --data_list_path data/train.csv python3 scripts/format_trials.py \ --voxceleb1_root your_voxceleb1_path \ --src_trials_path your_src_trail_path \ --dst_trials_path your_dst_trail_path |
语音数据默认为单声道wav文件,采用率为16k。 |
3. Training |
python3 tools/train.py \ --config your_yaml_path |
|
4. Log Visualization |
tensorboard --log_dir your_log_dir/ --bind_all |
|
5. Evaluation |
python3 tools/evaluate.py \ --config your_yaml_path \ --trial_path your_trial_path \ --checkpoint_path your_checkpoint_path |
|
6. Export |
语音声学特征的提取
由于 torchaudio 库中可能存在潜在的bug(经常各种warning),且未来我计划采用小波分析的方法(wavelet)进行一些其他的实验,因此提取声学体征的代码也是完全采用PyTorch手动实现(没有依赖于其他第三方库)。
Error code: 127
其中,值得注意的是:
- kaldi所实现的特征提取是offline的,采用PyTorch可以实现online的特征提取;
- Mel-Spectrogram 也叫Fbank,或是FilterBank,在Kaldi中叫Fbank中比较多,在TTS和VC中Mel-Spectrogram叫的比较多;
- 由我实现的特征提取代码,kaldi,librosa,torchaudio即便是在相同配置参数的情况下,所提取的到的结果都不同。
特征名 | 实现流程 | 代码存放位置 |
Spectrogram |
|
[link] |
Mel-Spectrogram |
|
[link] |
MFCC |
常用backbone与实现
backbone这个单词原意指的是人的脊梁骨,后来引申为支柱,核心的意思。在神经网络中,尤其是CV领域,一般先对图像进行特征提取(常见的有vggnet,resnet,谷歌的inception),这一部分是整个CV任务的根基,因为后续的下游任务都是基于提取出来的图像特征去做文章(比如分类,生成等等)。所以将这一部分网络结构称为backbone十分形象,仿佛是一个人站起来的支柱。
ResNet和其变种
ResNet
SENet
Res2Net
TDNN和其变种
TDNN(x-vector)
ECAPA-TDNN
ECAPA-TDNN architecture is based on the popular x-vector topology and it introduces several enhancements to create more robust speaker embeddings.
The pooling layer uses a channel-and context-dependent attention mechanism, which allows the network to attend different frames per channel. 1-dimensional SqueezeExcitation (SE) blocks rescale the channels of the intermediate frame-level feature maps to insert global context information in the locally operating convolutional blocks. Next, the integration of 1-dimensional Res2-blocks improves performance while simultaneously reducing the total parameter count by using grouped convolutions in a hierarchical way.
Finally, Multi-layer Feature Aggregation (MFA) merges complementary information before the statistics pooling by concatenating the final frame-level feature map with an intermediate feature maps of preceding layers.
The network is trained by optimizing the AAM-softmax loss on the speaker identities in the training corpus. The AAM-softmax is a powerful enhancement compared to the regular softmax loss in the context of fine-grained classification and verification problems. It directly optimizes the cosine distance between the speaker embeddings.
The model turned out to work amazingly well for speaker verification and speaker diarization.
VGG和其变种
VGG
RepVGG
常用pooling layer
Pooling Layer | ||
TSP | ||
TAP | ||
ASP | ||
SAP |
常用Loss
Loss | 描述 | 计算公式 | 代码实现 |
softmax | |||
Triplet Loss |
Triplet Loss基本思路是构造一个三元组,由anchor、positive 和 negative 组成,其中 anchor 和 positive 表示来自于同一个人的不同声音,negative 表示来自不同的人的声音,然后,用大量标注好的三元组作为网络输入,训练DNN参数。其优点在于直接使用embeddings之间的相似度作为优化的成本函数,最大化 anchor 和 positive的相似度,同时最小化 anchor和 negative 的相似度。这样,在提取了说话者的 embedding 之后,说话人识别任务就可以简单地通过相似度计算实现。 |
||
AM-softmax |
Kaldi搭建的声纹系统在模型训练中大多使用Softmax损失函数,但 Softmax 损失函数并不能增大类内紧凑性和类间分离性,为了增强embedding的判别性 |
||
AAM-softmax |
后端打分
打分方法 | 功能描述 | 计算公式 | 代码实现 |
cosine | |||
GPLDA | |||
LDA-GPLDA |
训练 trick 相关
tricks | 注意事项 |
distributed modes |
|
warm-up | |
learning rate schedule | |
optimizer | 尽管momentum SDG收敛速度慢于Adam,但是SGD的效果在实验上可能往往好于Adam(这个地方存疑),但在数据量大+做了数据增强后,使用SGD的训练时间实在是太久,综合来看还是使用Adam比较好 |
chunk size |
chunk size 对模型的训练效果有着非常显著的影响,最好把chunk size改小一点,不要太大。从目前来的实验结果看2秒往往优于2.5秒和3秒,不清楚chunk size更小能不能有更大的提升。 |
数据增强 | 数据增强有两种实现思路:
第一种直觉上可能会比第二种会更好,但实际使用过程中,第一种方法loss收敛速度很慢,我猜测可能在一个batch中做不同的数据增强会导致loss存在一定程度的发散。第二种还在开发中。 如果只是想快速验证一些科研想法,推荐不要开启数据增强。目前有文献表明,如果开启数据增强相比于不开启将会花费6倍的训练时间,数据增强应该是在
以上几种情况下使用 |
data loader 相关
数据增强
WavAugment |
|
SpecAugment |
|
训练数据类别均衡
在自然情况下,数据往往都会呈现长尾分布。这种趋势同样出现在从自然科学到社会科学的各个领域各个问题中,直接利用长尾数据来训练的分类和识别系统,往往会对头部数据过拟合,从而在预测时忽略尾部的类别。说话人数据类别分布不均衡可能会导致训练的模型效果一般,为此需要在dataloader上使用一些预处理的手段实现说话人类别的均衡。
长尾分布的最简单的两类基本方法是:
- 重采样(re-sampling):对头部类别的欠采样(under-sampling)和对尾部类别的过采样(over-sampling)
- 重加权(re-weighting):重加权则主要体现在分类的loss上
这两类方法本质都是利用已知的数据集分布,在学习过程中对数据分布进行反向加权,强化尾部类别的学习,抵消长尾效应。在这个项目中,我采用重采样(re-sampling)解决数据分布不均衡的问题。
说话人识别常用的数据集
评价指标相关
混淆矩阵
对于一个二分类系统,将实例分为正类(Positive)、负类(Negative),则模式分类器有四种分类结果:
- TP(True Positive):正确的正例,一个实例是正类并且也被判定成正类;
- FN(False Negative):错误的反例,漏报,本为正类但判定为假类;
- FP(False Positive):错误的正例,误报,本为假类但判定为正类;
- TN(True Negative):正确的反例,一个实例是假类并且也被判定成假类。
ROC 曲线
对于Speaker Verification任务,评估模型的方法可以是绘制ROC曲线(Receiver Operating Characteristic curve),首先了解以下定义
- 真阳率(True Positive Rate, TPR):描述识别出的所有正例占所有正例的比例
- 假阳率(False Positive Rate, FPR):描述将负例识别为正例的情况占所有负例的比例
- 真阴率(True Negative Rate,TNR):描述识别出的负例占所有负例的比例
- 其中其中TPR即为敏感度(sensitivity),TNR即为特异度(specificity)
ROC曲线将TPR定义为X轴,将FPR定义为Y轴; AUC(Area Under Curve)是指ROC曲线下面积,越接近1表示分类器越好。曲线下面积越大,分类的准确性就越高; 最靠近坐标图左上方的点为灵敏性和特异性均较高的临界值。 ROC曲线有个很好的特性:当测试集中的正负样本的分布变化的时候,ROC曲线能够保持不变。
Equal Error Rate(EER)
- 虚警率(False alarm)表示负类样本被分为正类样本在所有负类样本中的比例
- 漏警率表示(Missing alarm)表示正类样本被分为负类样本在所有正类样本中的比例
Minimum Detection Cost Function(minDCF)
Diarization Error Rate (DER)
Anti-Spoofing
对抗样本攻击与防御
当AI模型/算法设计之初未考虑相关的安全威胁的情况下,AI算法的判断结果容易被恶意攻击者影响,导致AI系统判断失准。其中最主要的安全威胁是闪避攻击,即是指通过修改输人,让AI模型无法对其正确识别。研究表明深度学习系统容易受到精心设计的输人样本的影响,这些输人样本称为对抗样本(Adversarial Examples)
说话人识别系统,无论是基于深度神经网络的/或是基于传统统计模型的i-vector系统都同样存在这样的问题。
对抗样本攻击
白盒攻击
攻击方法 | 计算方法 | 代码实现 |
BIM | ||
PGD |
黑盒攻击
对抗样本防御
工具代码和脚本
工具代脚本
脚本名称 | 实现思路与流程 | 代码位置 |
读取waveform |
目前各类开源的工具中,语音数据的读取的方法实现主要有两种:
在本项目中,根据training和evaluation阶段的不同,对语音的读取策略也有所区别。 |
[link] |
读取超参数 | 超参数的读入参考了nanodet项目的实现,采用了yacs来实现对yaml文件超参数对读取。 | [link] |
Voice Activity Detection(VAD) | VAD采用PyWebrct实现Python多进程处理 | |
信噪比(SNR)计算 | ||
准确率(Accuracy)计算 | ||
插值(interpolate) | ||
文档生成 | ||
format trials |
可视化
功能 | 效果预览图 | 代码链接 |
绘制语谱图(spectrogram) | ||
绘制3D语谱图(3D-spectrogram) | ||
绘制ROC曲线 | 生成缩略图出错:/bin/bash: /usr/bin/convert: No such file or directory Error code: 127 |
|
绘制PR曲线 | ||
绘制混淆矩阵 |
模型部署
以 PyTorch 和 TensorFlow 为代表的深度学习框架集成了模型的训练和推理两个过程。然而在实际模型的使用中,如果想要在不同类型的平台(云/Edge、CPU/GPU 等)上获得最佳性能,需需要调整模型(量化、知识蒸馏)和使用专门的推理库(ONNX,TensorRT)。
ONNX runtime 是一种用于将 ONNX 模型部署到生产环境的高性能推理引擎。 它针对云和 Edge 进行了优化,适用于 Linux、Windows 和 Mac。 它使用 C++ 编写,还包含 C、Python、C#、Java 和 Javascript (Node.js) API,可在各种环境中使用。 ONNX 运行时同时支持 DNN 和传统 ML 模型,并与不同硬件上的加速器(例如,NVidia GPU 上的 TensorRT、Intel 处理器上的 OpenVINO、Windows 上的 DirectML 等)集成。 通过使用 ONNX 运行时,可以从大量的生产级优化、测试和不断改进中受益。
代码规范
MISC
框架 | 描述 |
PyTorch |
PyTorch is a Python package that provides two high-level features:
|
PyTorch Lightning |
The goal of PyTorch Lightning is "You do the research. Lightning will do everything else". PyTorch Lightning was started by William Falcon while completing his Ph.D. AI research at NYU CILVR and Facebook AI Research, with the vision of making it a foundational part of everyone’s deep learning research code. The framework was designed for professional and academic researchers working in AI, making state of the art AI research techniques, such as TPU training, trivial. |
ONNX |
Open Neural Network Exchange (ONNX) is an open ecosystem that empowers AI developers to choose the right tools as their project evolves. ONNX provides an open source format for AI models, both deep learning and traditional ML. It defines an extensible computation graph model, as well as definitions of built-in operators and standard data types. |
NCNN |
ncnn is a high-performance neural network inference computing framework optimized for mobile platforms. ncnn is deeply considerate about deployment and uses on mobile phones from the beginning of design. ncnn does not have third party dependencies. it is cross-platform, and runs faster than all known open source frameworks on mobile phone cpu. Developers can easily deploy deep learning algorithm models to the mobile platform by using efficient ncnn implementation, create intelligent APPs, and bring the artificial intelligence to your fingertips. ncnn is currently being used in many Tencent applications, such as QQ, Qzone, WeChat, Pitu and so on. |
YACS |
YACS was created as a lightweight library to define and manage system configurations, such as those commonly found in software designed for scientific experimentation. These "configurations" typically cover concepts like hyperparameters used in training a machine learning model or configurable model hyperparameters, such as the depth of a convolutional neural network. |
Sphinx | Sphinx is a tool that makes it easy to create intelligent and beautiful documentation, written by Georg Brandl and licensed under the BSD license. |