DIGITAL SPEECH PROCESSING IN THE CONTEXT OF A HUMAN AUDITORY MODEL (SPEECH PROCESSING, AUDITORY MODEL)
Digital speech processing in the context of a digital hearing model can improve the subjective quality of the speech processing algorithms. This subjective quality is a measure of how 'good' the processed speech sounds to the listener. Subjective quality can be measured by paired comparison tests where subjects are asked to choose between two stimuli the one that sounds the best. This dissertation proposes that if digital speech processing is performed on speech that has been preprocessed using a digital hearing model, the resulting speech, after undoing the preprocessing of the digital hearing model, will sound 'better' as measured by subjective quality evaluations, whether or not standard objective distortion measures indicate otherwise. This dissertation proposes a digital hearing model for application in digital speech processing. This hearing model approximates the perception of intensity. Two digital processing algorithms were used to validate the claims of this dissertation. The first was spectral subtraction and the second was subband vector quantization. The results obtained from the subjective quality evaluations demonstrated evidence in support of the hypothesis of this dissertation. There was a 90% preference for coding in the perceptual domain for magnitude compression of 58:1 through 176:1 and a preference above 70% for noise suppression of speech corrupted by additive Gaussian noise of 18 dB and lower signal-to-noise ratios.