Feature Extraction and Dimensionality Reduction using IPS for Isolated Tamil Words Speech Recognizer

Abstract: Automatic Speech Recognition (ASR), is the process of converting a speech waveform into the text quite similar to the information being communicated by the speaker. This paper aims to construct a speech recognition system for Tamil language. Mel Frequency Cepstral Coefficients (MFCC) is a commonly used feature extraction technique for speech recognition which is computed by applying DCT to the mel-scale filter bank output. Instead of DCT, Integrated Phoneme Subspace (IPS) method is used to improve speech recognition. The experimental results show that the recognition accuracy of ASR using IPS in various forms yields higher or similar output comparative to MFCC and the word accuracy of one such form of IPS (IPS-2) is 84.00%.

Keywords: Linear transformation method, Hidden Markov Tool Kit (HTK), cepstrum, Hidden Markov Model (HMM).