Joint system for speech separation from speaking and non-speaking background, and de-reverberation: Application on real-world recordings

摘要

In real-life environment, the speech of interest is often correlated with different kinds of perturbation. Perturbation can be caused by speaking or non-speaking noise, or even by reverberation. This could make the speech signal auditable but not intelligible. In this case, speech cannot be exploited by other automated applications such as voice-command or speech/speaker identification and identification. Extracting a meaningful signal of good quality is a bigger challenge in monaural case. In this paper, we propose an extensible full joint system that deals with real-environment perturbations that include speaking and non-speaking background as well as reverberation. After introducing the input signal, a decision is taken on the process to opt for. The main chain blocks of the proposed system are speech denoising, speech separation and speech de-reverberation. The system operates in single channel case in a fully unsupervised manner. Furthermore, it requires minimal information about the reference signal. As the system is targeting real-time enhancement, results evaluation is conduct in terms of non-intrusive metrics in addition to intrusive metrics. The evaluation results prove the effectiveness of the proposed system in cancelling difficult noise types, in extracting desired speaker from speaking background, and in enhancing reverberated speech.