Speaker-Independent Isolated Word Recognition Using a Feed-Forward Artificial Neural Network
Project Award Date: 0000-00-00
In recent years Hidden Markov Model (HMM) and Dynamic Time Warping (DTW) algorithm have been successfully used in a developing a high performance speaker-independent isolated word recognition system. Even though HMM based word recognition system can achieve high recognition rates, the a-priori choice of the model topology hinders the flexibility of the model by including non-explicit knowledge of the speech recognition process. Also, testing a word using DTW algorithm needs extensive computing power if a real-time performance is required. On the other hand recent studies show that an artificial neural network (ANN) can be trained without the knowledge of the a-priori distribution of the data and the testing process is much faster than the DTW based algorithm. In this investigation, a feed-forward artificial neural network is trained using back-propagation learning algorithm to perform as a speaker independent isolated word recognizer. During training, the LPC coefficients are extracted from several repetitions of each word and these LPC coefficients are used at the input of the network as the feature vectors. An ANN is first trained, using back-propagation learning algorithm, to recognize ten different words. The network is then tested using the feature vectors for the ten words that are not used during the training process. The success rate achieved by the ANN is compared with the success rate achieved the DTW algorithm. The result indicates that a properly trained ANN can provide superior performance in word recognition as compared to the DTW algorithm, especially in a noisy environment.