Robust text independent closed set speaker identification systems and their evaluation

Abstract

This thesis focuses upon text independent closed set speaker identi cation. The contributions relate to evaluation studies in the presence of various types of noise and handset e ects. Extensive evaluations are performed on four databases. The rst contribution is in the context of the use of the Gaussian Mixture Model-Universal Background Model (GMM-UBM) with original speech recordings from only the TIMIT database. Four main simulations for Speaker Identi cation Accuracy (SIA) are presented including di erent fusion strategies: Late fusion (score based), early fusion (feature based) and early-late fusion (combination of feature and score based), late fusion using concatenated static and dynamic features (features with temporal derivatives such as rst order derivative delta and second order derivative delta-delta features, namely acceleration features), and nally fusion of statistically independent normalized scores. The second contribution is again based on the GMM-UBM approach. Comprehensive evaluations of the e ect of Additive White Gaussian Noise (AWGN), and Non-Stationary Noise (NSN) (with and without a G.712 type handset) upon identi cation performance are undertaken. In particular, three NSN types with varying Signal to Noise Ratios (SNRs) were tested corresponding to: street tra c, a bus interior and a crowded talking environment. The performance evaluation also considered the e ect of late fusion techniques based on score fusion, namely mean, maximum, and linear weighted sum fusion. The databases employed were: TIMIT, SITW, and NIST 2008; and 120 speakers were selected from each database to yield 3,600 speech utterances. The third contribution is based on the use of the I-vector, four combinations of I-vectors with 100 and 200 dimensions were employed. Then, various fusion techniques using maximum, mean, weighted sum and cumulative fusion with the same I-vector dimension were used to improve the SIA. Similarly, both interleaving and concatenated I-vector fusion were exploited to produce 200 and 400 I-vector dimensions. The system was evaluated with four di erent databases using 120 speakers from each database. TIMIT, SITW and NIST 2008 databases were evaluated for various types of NSN namely, street-tra c NSN, bus-interior NSN and crowd talking NSN; and the G.712 type handset at 16 kHz was also applied. As recommendations from the study in terms of the GMM-UBM approach, mean fusion is found to yield overall best performance in terms of the SIA with noisy speech, whereas linear weighted sum fusion is overall best for original database recordings. However, in the I-vector approach the best SIA was obtained from the weighted sum and the concatenated fusion.