Improving Statistical Model-Based Speech Enhancement with Deep Neural Networks

Enhancing and cleaning speech signals in noisy environments has long been a challenge in the field of electronic communication and audio processing. To improve the quality and intelligibility of speech, effective single-channel speech enhancement techniques are desperately needed. The problem becomes even more critical when the input signal is deteriorated by various types of noise and reverberation. The current approaches rely heavily on explicit statistical models for noise and speech, often leading to suboptimal performance caused by inaccurate noise estimation and insufficient suppression of reverberations. These current techniques struggle in dynamic and reverberant environments, hindering their ability to effectively separate speech from noise and leading to poor speech quality and decreased intelligibility.

Technology Description

The invention is a technology that uses deep neural networks (DNN) to improve the performance of single-channel speech enhancement systems. Embodiments feature a DNN-trained system capable of predicting the presence of speech in an input signal, along with a framework for tracking ambient noise and estimating the signal-to-noise ratio. This system offers increased flexibility in its design parameters, like gain estimation, and enables joint suppression of both additive noise and reverberation. What sets this technology apart is its ability to function effectively in the presence of both noise and reverberation. Designed to detect speech amidst noise and reverberation, the system has a leading edge in comparison to baseline systems. Furthermore, it is capable of significant improvements in objective speech-quality metrics, thereby outperforming traditional baseline systems.