System and method for speech enhancement using dynamic noise profile estimation

Abstract

A speech-enhancing noise filter is disclosed. The noise filter comprises a microphone for acquiring speech data from a user; a feature extraction module configured to extract a plurality of features characterizing the speech data; a neural network configured to receive the plurality of extracted features and to estimate a noise profile from the plurality of extracted features; a noise removal module configured to remove the noise profile from the noisy speech data; and a reconstruction module configured to generate a waveform from the plurality of frames after removal of the noise profile from each of those frames. The neural network is trained to isolate various types of noise from the user speech in the speech data and then subtract the noise from the speech data, thus leaving only the user speech free of noise. This filtering is dynamically performed on a frame-by-frame basis from each frame of the speech data, thereby making it possible to specifically identify and remove different types and levels of noise in each frame. The noise filter of the present invention accurately removes noise independent of the language in which the user utters the speech data.