DNN与谱减法相结合的语音分离技术

Speech Separation Combined with DNN and Spectral Subtraction

针对传统DNN语音分离中噪声干扰的问题,提出了一种在DNN语音分离后期处理中结合DNN和谱减法的语音分离方法。首先提取语音声级特征,通过DNN学习带噪特征到分离目标语音的映射,得到分离目标语音;然后对分离目标语音中每一时频单元进行噪声能量估计;最后,通过快速傅里叶逆变换得到谱减后的分离语音波形信号。通过对不同类型的噪声和不同输入信噪比混合后的语音信号进行试验,结果表明,加入谱减法后分离的语音信号与只经DNN网络输出的语音信号相比,前者分离的语音可懂度和信噪比得到了显著提高,并且分离语音的信号更接近于纯净语音的信号。

In view of the problem of noise interference in traditional DNN speech separation,a speech separation method based on DNN and spectral subtraction was proposed in the post processing of DNN speech separation.Firstly,the features of speech were extracted and the DNN was used to learn the mapping of the noisy features to the separated target speech.Then the noise energy is estimated for each time frequency unit in the separated target speech.Finally,the speech waveform was obtained by the inverse fast fourier transform.By testing the speech signal mixed by different types of noise and different input SNR,the experimental results show that compared with the speech signal output only by the DNN network,the speech signal separated after adding spectral subtraction is significantly improved in the speech intelligibility and signal to noise ratio of the proposed algorithm.The similarities between the separated speech signal and the original clean speech signal has also been greatly improved.