Nonlinear speech processing with oscillatory neural networks for speaker segregation

Nonlinear masking of space-time representations of speech is a universal technique for speech processing. In the present work we use an AM representation of cochlear filterbank signals in combination with a mask that is derived from a network of oscillatory neurons. The proposed approach does not need any training or learning and the mask takes into account the dependence between points from the auditory derived representation. A potential application is illustrated in the context of speaker segregation.

2 Introduction

Speech enhancement, speaker segregation, speech recognition and coding can be viewed as the result of 3 successive processes. 1

Decomposition of the speech signal into an adequate space-time representation (auditory image representations, spectrograms, wavelet decompositions, etc.);

2

Selection of the relevant information from the chosen representation or masking of the irrelevant information;

3

Synthesis of the speech (enhancement, speaker segregation, coding) or extraction of the parameters from the selected relevant areas of the representation (speech recognition).