DNN-based Speech Enhancement with Harmonics Regeneration
* Präsentierender Autor
Zusammenfassung:In a hands-free telephony setting for automotive environments, driving noise distorts both the phase and amplitude of acquired speech signals. The aim of noise reduction approaches is to restore clean speech before sending it over the telephone channel. State-of-the-art approaches use DNN-based filters that estimate real or complex–valued filter coefficients in the frequency domain. While real-valued filters can only correct the spectral magnitude, complex-valued filters can also correct the phase. But the latter comes at a significantly higher computational cost.In this work, we take an alternative approach that re-synthesizes the harmonic structure of speech by using DNN-based pitch trackers and voiced/unvoiced detectors. The re-synthesis concentrates on the frequency band between 0 and 1000Hz as the human hearing system is most sensitive to phase errors in this range. Higher frequency bands of the speech signal are obtained with a real-valued filter whose coefficients are estimated with a DNN.The perceived speech quality of the processed speech is evaluated in subjective listening tests.