Single Microphone Speech Noise Separation

Abstract

The technique of extracting the target speech from a noisy speech mixture collected using a single microphone is known as monaural (single microphone) speech separation. Mobile telephony, hearing aid design, and robust automatic speech and speaker identification are only a few of the applications. Researchers recently succeeded in separating the clean speech from the monaural noisy speech mixture using the computational auditory scene analysis (CASA) methodology. The spectral magnitude mask (SMM) has been presented as a computational aim to increase speech intelligibility and speech quality in CASA-based single microphone speech noise separation approaches. The power spectrum, Mel-frequency cepstrum coefficient (MFCC), and Relative spectral perceptual linear prediction (RASTA-PLP) properties of the noisy mixture are discussed in this study, as well as short-time Fourier transform spectral magnitude masking as a training target. These attributes are fed into DNN, which serves as a distinct clean speech target. The performance of separated speech is assessed using STOI and PESQ.