Effects of time lapse on Speaker Recognition results

Abstract

The effect of time lapse has not been studied well in most biometrics. Here, this effect is studied for Speaker Recognition, namely, Speaker Identification and Speaker Verification. The RecoMadeEasyTM speaker recognition engine has been used to obtain baseline results for 22 speakers who have been involved in a long-term study. The speakers have given data in three seatings with 1 to 2 months delay between consecutive collections. The speakers were real proficiency test candidates who were asked to speak in response to prompts. At each seating, several recordings were made in response to different prompts. The error rates are discussed, going from one seating to the next, for Identification and Verification. Large degradations are seen across different seatings. Two different adaptation techniques have been studied for reducing this discrepancy with very promising results.