Speaker Verification in the Presence of Channel Mismatch Using Gaussian Mixture Models
AIR FORCE INST OF TECH WRIGHT-PATTERSONAFB OH SCHOOL OF ENGINEERING
Pagination or Media Count:
A channel compensation method is sought for use in speaker identification ID and verification applications under matched and mismatched training and testing conditions. This work expands on previous work on matched conditions by investigating three techniques on matched and mismatched conditions using the TIMIT and NTIMIT speech databases. First, previous results on 168 speakers are reproduced for matched conditions using Gaussian mixture models GMM and mel-frequency cepstral coefficients. Next, cepstral mean subtraction with band limiting CMSBL is investigated. The third method, developed in this thesis, uses a modified Wiener filtering approach to channel compensation. New GMMs are created for each method. The first approach is then expanded to include all 630 TIMIT and NTIMIT speakers for speaker verification. For speaker ID under matched conditions, the CMSBL method had three more errors than no additional preprocessing but yielded the best ID results for the mismatch case with 27.4 correct. Additionally, the CMSBL method yielded the best verification results with an equal error rate of approximately 0.26 for matched conditions on TIMIT and approximately 19.6 for mismatched conditions on NTIMIT.