A multi-class SVM based phonemes classifier based on a trainable confidence measure

Although the recognition results of support vector machines(SVM) are very promising in many applications, there is a gap between the accuracy of SVM based speech recognizers and time series models (e.g. hidden Markov model) in speech recognition. The main reasons are the lack of proper methods to classify the acoustic units into more than two classes and suitable SVM based sequence decoders. This paper describes a trainable method for SVM multi-class classification based on confidence measures of the sets of two-class SVM classifiers using an artificial neural network. In addition, a pruning method has been proposed for SVM multi-class classification to decrease the computational complexity without significant decrease in accuracy. Also, a method has been proposed for time series recognition of the feature vectors of an utterance based on SVM classifiers. The experiments have been conducted on a set of confusable phonemes using TIMIT corpus. The results of the first method show 10% and 6% relative improvements in the recognition rate in comparison to one-versus-one and Kruger methods respectively for /b/, /d/ and /g/ phonemes. In addition, it is empirically deduced that the proposed phoneme classification framework yields significantly better classification rates than classic voting method. Comparing phoneme classification results of the proposed method with one-versus-one method indicate a 26% improvement in the classification rate for /b/, /d/ and /g/ phonemes.