New features using fuzzy c-means alogorithm for automatic language recognition

We propose new features for the language recognition using Gaussian computations. New features are derived from traditional features like Mel frequency cepstral coefficients (MFCC) using fuzzy c-means clustering algorithm. MFCC feature vectors derived from huge corpus of all languages under consideration are grouped into c-clusters using fuzzy c-means clustering algorithm and one Gaussian distribution is modeled for each cluster. In the training phase, new feature vectors are derived from language specific speech corpus using the clusters which are formed by fuzzy c-means clustering algorithm. In the testing phase, similar procedure is followed for the extraction of c-element feature vectors from unknown speech utterance, using the same c-Gaussians and evaluated against language specific HMMs. The language apriori knowledge (usefulness of feature vector) has been considered for the improvement of recognition performance. Continuous hidden Markov model (CHMM) is designed using the new feature. The languages in OGI database are used for the study and we have achieved good performance.