Designing a fast neuro-fuzzy system for speech noise cancellation

Abstract

Noise canceling is an adaptive interference filtering technique that has shown to be highly adventageous in many applications where fixed filters are not efficient. We present an experimental neuro-fuzzy inference system, based on the ANFIS architecture, which has been implemented with the objective to perform nonlinear adaptive noise cancellation from speech. The novelty of the system described in the present paper, with respect to our previus work, consists in a different set up, which requires two inputs with seven membership functions each, and uses a second order sinc function to generate the nonlinear distortion of the noise. This set up allows a better generalization to the system for learning the noise features. Indeed, the system was trained only once during few epochs, with a sample of babble noise, but it was able to clean speech sentences corrupted not only with the same noise, but also with car, traffic, and white noise. The average improvement, in terms of SNR, was 37 dB without further training, resulting in a great reduction of the computational time. © Springer-Verlag Berlin Heidelberg 2000.