Human and machine performance on speaker identity verification

Abstract

Two experiments were conducted to identify acoustic features for speaker identity verification (SIV) that are used by humans and not by cepstral-based algorithms. Although these algorithms generally out-perform human listeners for randomly selected comparisons between single-word utterances, this approach was to analyze human performance on comparisons that could not be effectively discriminated by machine. Experiment 1 showed that humans could perform at high levels of accuracy on these comparisons suggesting that either information exists that is not captured by the algorithms, or that the information is coded by the algorithms but is not used effectively. The second experiment consisted of three stimulus conditions for SIV; digitized speech signals, noise-excited resynthesized LPC signals, and error prediction signals from the LPC. Results indicated high levels of performance in the natural and error prediction signal conditions and performance near chance in the noise excited condition, thus suggesting that the error signal provides valuable information that allows humans to distinguish between speakers. It may be possible to improve verification algorithms by adapting current models to more accurately utilize information used by human listeners.