A real-time kepstrum approach to speech enhancement and noise cancellation

Abstract

This paper introduces the kepstrum method of identification of acoustic transfer functions and its real-time application in techniques such as speech enhancement and noise cancellation. By using kepstrum (known elsewhere as complex cepstrum) analysis, we will show that the kepstrum method is reliable and applicable to real-time processing by providing an efficient kepstrum processing technique with a fast processing time using fast Fourier-transforms (FFTs). Furthermore, it will be shown that its application to speech enhancement methods gives several favourable effects such as: (1) application flexibility by enabling invertibility, which gives a stable minimum phase transfer function on application of the inverse of a room impulse response, (2) processing robustness by using it as building blocks and (3) computational simplicity by highly reduced weights in an adaptive algorithm when it is used with a cascaded adaptive filter. Accordingly, we will show that the kepstrum approach contributes to speech enhancement and noise cancellation with an improved performance in signal-to-noise ratio (SNR) by using only two microphones with a small physical dimension in size. Experiments on a 20 cm broadside two-microphone acoustic beamformer configuration are implemented in real-time in an indoor office.