Neural Network-based Voice Recognition Robustness

Although speech recognition products are already available in the market at present,their development is mainly based on statistical techniques which work under veryspecific assumptions, most speech recognition systems are still in their infancy and haveproblems if migrated from laboratory to actual applications. Aiming at the robustness ofspeech recognition system, this dissertation attempts to study, in depth, on the theory andtechniques of speech recognition by using the concerned experiments.The structure of the dissertation is as follows: The first chapter gives the researchbackground with some known results. The second chapter introduces the basic conceptsof this domain. In third chapter, we discuss the feature extraction of the speech signal.The model based the neural network is in forth chapter. The last chapter is theconclusion.Our research is show as follows:A speech recognizer system comprised of two distinct blocks, a feature extractorand a recognizer. The feature extractor block uses a standard LPC (Linear PredictiveCoding) Cepstrum coder, which translates the incoming speech into a trajectory in theLPC Cepstrum feature space. According as the characteristic of the short-time averageenergy and short time zero-crossing in the speech signal, we discuss the method of thestarting-point decision, also design and implement the voicing decision of the speechsignal using the wavelet transform. The pitch detection and formant detection in themandarin digit speech recognition show a high performance. Noise robustness andemotional feature analysis become the hardness of speech signal processing.Designs of the recognizer blocks based on three different approaches are compared.The performance of Multi-Layer Perceptrons, Time Delay Neural Network and RecurrentNeural Network based recognizers will be discussed.