Evaluation of optimal and sub-optimal speech noise reduction wiener filters

This paper presents a comparative evaluation of Wiener optimal and sub-optimal finite impulse response filters, which allows a balance between noise reduction and distortion insertion, by setting a parameter α. This is done observing an Automatic Speech Recognition (ASR) system error rate. The ASR system in these paper is tested for Brazilian Portuguese. The percentage of correctly recognized words is obtained for speech signals subject to Additive White Gaussian Noise (AWGN), for an SNR ranging from zero to 20 dB, using filtered speech signals. To evaluate the distortion effect caused by filtering, the filtered version of clean speech signals is processed by the recognizer, and it is observed that the error rate decreases with the reduction of the parameter α. The application of a suboptimal filter, with α = 0.7, produces the highest recognition rate. The observed improvement is 10% for the lowest SNR and 14% for the highest SNR. It is observed that the output SNR increases with parameter α.