Comparison of Three Methods of Eliminating Musical Tones in Speech Denoising Subtractive Techniques

Speech enhancement techniques, using spectral subtraction, have the drawback of generating an annoying residual noise with musical character. In this paper, we use and compare different post-processing methods of musical tones reduction. All these approaches need both a musical noise detector and a musical noise reducer. For each part, we propose ´non perceptual´ and perceptual based approaches. The musical noise detector is based either on shape characteristics or masking consideration. To reduce musical noise, three techniques are used. The first eliminates musical tones in time domain by adding a synthesized signal to the denoised signal. The second operates in the spectral domain and consists in shifting down the detected musical tones under a proposed masking threshold. The third consists in shape modification of detected musical tones. Experiments results show that perceptual based approaches are the best ones.