![]() | Only 14 pages are availabe for public view |
Abstract The thesis introduces a newly designed feature extraction method for speech signal that can be used in Automatic Speech Recognition (ASR). The newly developed methods are inherited from the wavelet packets decomposition of speech signal. The work is an enhancement of the original features developed earlier by other researchers. Automatic Speech Recognition (ASR) of Arabic Phones without Grammar is considered as a problem to be solved. Information related to speech phoneme is encoded into 15 bits instead of 7 bits in the original version of Best Tree Encoding (BTE4). Best Tree Encoding of 5 levels of wavelet analysis (BTE5) gives 25% efficiency enhancement over the original BTE4 for solving ASR problem. Whenever possible the comparison results are provided to explain the trend of enhancement in the Success Rate (SR). In Addition; BTE5 achieved 25% SR while Mel- Frequency-Cepstral-Coefficients (MFCC) achieved 39% Success Rate of the ASR problem. This results in BTE5 achieving 64 % of the SR of the popular and famous (MFCC) for the same ASR problem |