On Factors Affecting MFCC-Based Speaker Recognition Accuracy

We evaluate the accuracy of an MFCC-based speaker recognition method. We analyse the recognition results using speech signal from everyday life environments. We study the mismatch effects of text-dependency, sample length, language, style of speaking, cheating, microphone, sample quality, and noise. The experiments on a self-collected corpus of 30 subjects indicate that any mismatch degrades recognition accuracy. The most dominating factors are noise, microphone, disguise, and degrading of the sample rate and quality. Speech-related factors and sample length are less critical.