Using the time frequency structure of pitch periods to improve speaker verification systems

Commonly used robust speaker verification systems are based on time-varying autoregressive spectral estimation (AR) combined with hidden Markov modeling (HMM) or dynamic time warping (DTW). An exhaustive optimization of these methods in the past has culminated in quite reliable verification schemes. It seems unlikely, though, that further significant improvements are readily obtained along the same path. Unlike time-varying AR-modeling, which focuses on the the global spectral structure of an utterance, we are introducing a new method that focuses on the local time-varying spectral structure of individual pitch periods. Additionally, a pattern classification method using singular value decomposition (SVD) is employed. The new method by itself does not deliver better results than commonly used global methods; however, it is shown that an acceptance/rejection decision derived from both global and local analysis greatly improves the performance of the verification system.