论文笔记:DEEP LEARNING FOR MONAURAL SPEECH SEPARATION
论文笔记:DEEP LEARNING FOR MONAURAL SPEECH SEPARATION
Po-Sen Huang, Minje Kim, Mark Hasegawa-Johnson, Paris Smaragdis
introduction
对语音信号的source separation,信源分离,在很多场合都有应用。比如把noise和speech分开可以提高ASR (automatic speech recognition)的精度。还有把歌声和音乐分开,等等。传统方法有NMF,即非负矩阵分解,和PLSI,probabilistic latent semantic indexing,它们是学习 t-f 域的表示,也就是学习非负数的重构的bases和weights。
In this paper, we explore the use of a DNN and the use of an RNN for monaural speech separation in a supervised setting. We propose the joint optimization of the network with a soft masking function. Moreover, a discriminative training
objective is also explored. The proposed framework is shown in Figure 1.
网络结构和流程图如上。可以看出,实际上是学习的语音信号的time-frequency的mask,然后再反变换到时域。用的网络就是RNN,学mask有两种,一种是hard,一种是soft。hard也就是binary的,非此即彼,soft就是学习一个比例的图,类似于heatmap一类的东西。如下:
Discriminative Training :
Empirically, the value γ is in the range of 0.05∼0.2 in order to achieve SIR improvements and maintain SAR and SDR.
因为评判指标里有SIR,signal to interference ratio,所以要求另外一个约束,也就是分出来的两个要尽量远离,也就是最大化类间差。
后面的内容涉及到对语音信号的测试和评估,略。
2018年06月06日22:09:18
有一个可以想念的人,就是幸福。 —— 导演,岩井俊二