Abstract
In processing voice with environment noise, the noise must be eliminated to improve the vocabulary recognition rate. In this process, noise elimination and feature extraction for model-estimate technologies are utilized. Concerning these noise-elimination and model-estimate technologies, the most important part is to estimate mixed noise in the source signal and eliminate it. In a vocabulary recognition system, if unexpected noise appears in the signal, or if quantization noise is basically added to digital signals, the source signal is changed or damaged, which decreases the recognition rate. If a source signal is transformed or changed by being mixed with diverse kinds of noise, the hidden Markov model (HMM) is used for effective noise elimination. The HMM forms a model by extracting features to flexibly respond to diverse vocabulary changes found in voice and text, etc. The method is applicable to data changing over time, and can establish a more effective model as the number of parameters constituting the model grows larger. It can provide a robust model estimate by using a parameter set for structured models. HMM-based vocabulary recognition shows discriminating distribution of recognition probability regarding recognition vocabulary models, and has lower computational complexity for recognition. But it produces a relatively lower recognition rate. To solve that problem, a vocabulary recognition-model optimization method is proposed based on a similar phoneme–recognition process and efficient feature extraction. In vocabulary recognition, a similar phoneme–recognition process is applied to HMM to recognize models adjacent to the model group. Efficient feature extraction is used to optimize the recognition model to enhance the recognition rate. For vocabulary composition, a Gaussian-mixture feature-extraction model is optimized and used as a vocabulary recognition model. Then, it is processed with similar-phoneme recognition regarding the vocabulary recognition model.