Speech Recognition System based on Discrete Wave Atoms Transform Partial Noisy Environment

Automatic speech recognition is one of the most active research areas as it offers a dynamic platform for human-machine interaction. The robustness of speech recognition systems is often degraded in real time applications, which are often accompanied by environmental noises. In this work, we have investigated the efficiency of combining wave atoms transform (WAT) with Mel-Frequency Cepstral Coefficients (MFCC) using Support Vector Machine (SVM) as classifier in different noisy conditions. A full experimental evaluation of the proposed model has been conducted using Arabic speech database (ARADIGIT) and corrupted with "NOISEUS database" noises at different levels of SNR ranging from -5 to 15dB. The results of Simulation have indicated that the proposed algorithm has improved the recognition rate (99.9%) at 15 dB of SNR. A comparative study was conducted by applying the proposed WAT-MFCC features to multilayer perceptron (MLP) and hidden Markov model (HMM) in order to prove the efficiency and the robustness of the proposed system.