One way to keep up a decent recognition of results- with increasing vocabulary- is the use of base units rather than words. This paper presents a Continuous Speech Large Vocabulary Recognition System-for Arabic, which is based on tri-phones. In order to train and test the system, a dictionary and a 39-dimensional Mel Frequency Cepstrum Coefficient (MFCC) feature vector was computed. The computations involve: Hamming Window, Fourier Transformation, Average Spectral Value (ASV), Logarithm of ASV, Normalized Energy, as well as, the first and second order time derivatives of 13-coefficients. A combination of a Hidden Markov Model and a Neural Network Approach was used in order to model the basic temporal nature of the speech signal. The results obtained by testing the recognizer system with 7841 tri-phones. 13-coefficients indicate accuracy level of 58%. 39-coeefficents indicates 62%. With Cepstrum Mean Normalization, there is an indication of 71%. With these small available data-only 620 sentences-these results are very encouraging.
展开▼
