Robust speaker identification using multi-microphone systems

A dual microphone system to improve the accuracy of speaker identification in adverse conditions such as noisy background environments is proposed. The extra information obtained from the second microphone is used to deemphasize the speech contamination of the noise estimate. This noise estimate is then used in a spectral subtraction procedure which provides a more accurate estimate of the speech signal. Results indicate that the length-normalised log likelihood of test sequences is better than the log likelihood of the original two signals.