“Self Learning Speech Recognition Model Using Vector Quantization” Abstract:
In this project we have designed a “Self Learning Speech Recognition Model Using Vector Quantization” technique for modeling. The project has two phases namely, the recognition phase and the self learning phase. Initially reference models of different words are created by giving the input speech samples. The input speech sample passes through various steps of speech recognition system and the feature vectors are generated from each speech sample and are stored in the reference models. Now in the recognition phase a test sample is given to the system. After passing through various steps some feature vectors are extracted from this test sample. These feature vectors are compared with the reference models one by one. The best matching reference model is shown as the recognized word. In the self learning phase the recognition process takes place first and then learning of reference model is performed. In this phase the training samples are given to the system. Again features are extracted. These features pass through the recognition process and a word is recognized. There may be a possibility that the recognition becomes incorrect due to high noise. Under such a case, if the recognition is inaccurate then we shift our speech recognition system phase to learning phase. Now our test sample becomes our training sample. The training sample is assigned an identity. This identity allows the sequential access of a specific reference model to which that identity belongs. After sequentially accessing the reference model feature vectors the training sample feature vectors are added to the reference model feature vectors through vector quantization under human supervision so that there remains no chance of addition of training sample in the wrong reference model. This property makes our system a self learning speech recognition system. With a constant learning of the speech...
Please join StudyMode to read the full document