Multiclass SVM based spoken Hindi numerals recognition

الملخص EN

This paper presents recognition of isolated Hindi numerals using multiclass Support Vector Machine (SVM).

The acoustic features in terms of Linear Predictive Coding (LPC), Mel-Frequency Cepstral Coefficients (MFCC) and combination of LPC and MFCC have been considered as inputs to the recognition process.

The extracted acoustic features are given as input to the SVM.

The classification is performed in two steps.

In first step, a one-versus-all SVM classifier is used to identify the Hindi language.

Further, in second step ten one-versus-all classifiers are used to recognize numerals.

The linear, polynomial and RBF kernels are used for the construction of SVM for recognition purpose.

In the first phase, the best kernel strategy was explored for a fixed number of frames of the speech signal.

The highest recognition rate has been achieved using linear kernel strategy.

Next, the number of frames in order to calculate LPCs and MFCCs was varied and recognition accuracy was calculated.

The highest recognition accuracy achieved in this study is 96.8%.