收录:
摘要:
In this paper, we present three strategies to achieve speech enhancement, which is based on Cepstral Mapping and Deep Neural Networks (DNN). Firstly, we apply DNN to directly predict the clean speech Cepstral feature given noisy Cepstral input. Then, by waveform reconstruction, we can obtain desired clean speech. Comparing with the method of directly mapping log-power spectral (LPS), our method is able to be more effective to recover speech harmonic structure and gain the higher speech quality. Additionally, we also utilize DNN to estimate ideal Wiener filter by giving noisy Cepstral input. Finally, a fusion framework is proposed to acquire enhanced speech signal, which combines Cepstral feature mapping and Wiener filter. Experiments show that the proposed algorithms are able to achieve the state-of-the-art performance in improving the quality and intelligibility of noisy speech. © 2018 IEEE.
关键词:
通讯作者信息:
电子邮件地址: