Performance Improvement of Continuous Speech Recognition System Using Extracted Features of Speech Manifolds in the Reconstructed Phase Space

Design of new feature extraction methods out of the speech signal and combination of their obtained information are the most effective approaches to improve the performance of automatic speech recognition (ASR) system. Recent researches have been shown that the speech signal contains nonlinear and chaotic properties, but the effects of these properties were not used in the continuous ASR systems. Reconstructed phase space (RPS) is an appropriate domain to exhibit nonlinear properties of a chaotic signal. Therefore, in this paper a new method is proposed to utilize the RPS-based features (LLRPS). These features will be computed using similarity scores between the embedded speech signal in the RPS and a set of predefined phoneme manifolds. Then, TMLP-based neural network estimates phoneme posterior probability over the LLRPS features. This network includes some useful properties such as extracting dynamic information and output combination methods. Experimental results using Farsdat speech database show that nonlinear combination of the speech recognition outputs including traditional MFCC features and LLRPS features, leading to improvement of 3.94% and 4.02% in the accuracy of frame and phoneme recognition, respectively.