A new way to enhance speech signal based on compressed sensing

We propose a novel speech enhancement approach based on compressed sensing. The method performs noise subtraction in the measurement domain in addition to sparse recovery. Dictionary learning, using K-singular value decomposition algorithm, is performed to create an overcomplete dictionary. The noise in the measurement domain is estimated during pauses. Voice activity detection (VAD) is used to classify speech/silence frames. Based on the VAD output, a mask function is created, and applied to the noisy speech spectrogram. Furthermore, from each active-speech observation vector, the estimated noise observation vector is subtracted. The enhanced speech spectra are obtained by sparse recovery using orthogonal matching pursuit. Our method is tested for various types of noise, including babble, market, police siren, piano, factory, and white noises. Comparison with recent state-of-the-art methods is performed in terms of segmental signal to noise ratio, perceptual evaluation of speech quality, and short-time objective intelligibility. The results reveal the advantages of the proposed method.