Power spectral analysis and Neural network for feature extraction and recognition of speech

Abstract

The research is about the feature extraction and recognition of Bangla (Bengali) phoneme, word and command. Power spectral analysis as a feature extraction technique used. Each frame of a speech signal (spectrum) divided into four parts to take the mean absolute value of each part as a pattern recognition technique. For pattern recognition and training purpose Neural network used for the recognition of Bangla phoneme, isolated word and command. The tools and techniques applied in the research provide good result in recognition for single male or female in Bangla phoneme, word and command. Speech recognition rate reduces when scale of experiment increases. Changing of window length during framing slightly influences recognition rate. Linear predictor coefficient analysis, Mel frequency cepstral coefficient could be taken into consideration for future research.