Deep Speaker: an End-to-End Neural Speaker Embedding System

实现流程:
Deep Speaker: an End-to-End Neural Speaker Embedding System

Front Processing:
语音输入被转换为64维fbank,并且含有零均值与单位方差。

DNN:有两种DNN:
* ResCNN
* GRU

ResCNN:
Deep Speaker: an End-to-End Neural Speaker Embedding System

GRU:
Deep Speaker: an End-to-End Neural Speaker Embedding System

Average Sentence:
Deep Speaker: an End-to-End Neural Speaker Embedding System

将帧级输入聚合为整段语音的输入

Affine:将其转换成512维的embedding。

计算相似度:
Deep Speaker: an End-to-End Neural Speaker Embedding System

最后用triplet loss为目标进行训练

实验
使用softmax和交叉熵损失来预训练整个模型,即用一个classification layer来代替length normalization和triplet loss层。
Deep Speaker: an End-to-End Neural Speaker Embedding System

可见与训练提高了模型准确率
Deep Speaker: an End-to-End Neural Speaker Embedding System

可见实验结果很好,且ResCNN效果好于GRU。
Deep Speaker: an End-to-End Neural Speaker Embedding System

将两种网络的score相加,可以看到表现得到了提升,其中score fusion表示把两个模型输出的cos score相加。
Deep Speaker: an End-to-End Neural Speaker Embedding System

在文本无关的说话人验证任务上,训练数据集越大,模型的训练越充分,效果越好
Deep Speaker: an End-to-End Neural Speaker Embedding System

deep speaker具有比较好的跨语言能力

上一篇:GRU反向传播公式推导


下一篇:一个基于RNN的实体抽取工具