The aim of the speaker recognition system is identifying who is speaking, by the personal identity information extracted from speech signal. Practical application mainly used for speaker recognition is authentication of a person on the telephone line. But the telephone speech contains nonlinear distortions caused by transmission line, which can lead to serious performance degradation due to the mismatches between training and testing environments. Some compensation methods such as CMS (cepstral mean subtraction) and SBR (signal bias removal) were proposed. But these have their own limits on the estimation of time-varying channel distortions and the preservation of static speaker information while compensating the distortions, so new method suited to speaker recognition is needed.
This thesis proposes feature parameter transformation using ICA (independent component analysis) as a new compensation method. ICA is a signal processing technique, whose goal is to express a set of random variables as linear combinations of components that are statistically as independent from each other as possible. The proposed method assumes that the cepstrum vectors from various channel-conditioned speech are linear combinations of some characteristic functions with random channel noise added, and transforms them into new vectors using ICA. The resultant vector space can give emphasis to the repetitive speech information and suppress the random channel distortions.
The proposed method was compared to other channel compensation methods. Experiments on SPIDRE, real telephone speech database, were performed in equal and different channel conditions. In the equal channel condition, proposed method marked 2%~7% higher recognition rate than others. In the different channel condition, 9%~16% higher. These results showed that the proposed method is more robust and discriminating than others.
