Speaker identification by computer and human evaluated on the SPIDRE corpus
Hassan Ezzaidi, Jean Rouat · Canadian acoustics · 2000
INTRODUCTION Although many experiments on clean speech report high identification rates for computer systems, results on noisy telephone speech with different handsets are usually too poor for practical identification tasks (noise, limited bandwidth, effect of the channel, telephone handsets variability)[1]. What would be the identification rate of humans in the same conditions? A reference is necessary in order to evaluate the performance of computer systems. The comparison between computer and human has been already made. For a review one can refer for example to the work by Doddingtion [2]. As the performance of human has been chown to be dependent of the speech nature, we propose to examine the effect of telephone handset variability for text-independent speaker identification of telephone speech. We report human and computer speaker identification with the SPIDRE database. Section 2 describes the experimental conditions while section 3 and 4 are the results and discussion. Sect