Study On Robust Feature Extraction Method Of Speech And Audio-based Context Recognition

Audio signal contains a rich of information, and has the advantages of non-contact, nature and low-cost of capture equipment, which makes audio-based recognition technology has a good prospect in the human-computer interaction and the intelligence of electronic products. In practical applications, speech signals are often disturbed by environments which reduce the speaker recognition accuracy. The robust feature extraction methods can effectively reduce affection of noises and improve the speaker's recognition accuracy. Research on robust feature extraction methods has become a popular area in audio-based recognition filed. Audio-based context recognition technology can make the computer automatically sense environmental characteristics and have auditory intelligence. Now, audio-based context recognition has gained more concern from researchers.Based on previous researchers' efforts and achievements, this thesis is mainly about robust feature extraction methods in speaker recognition, and features selection methods, model optimization method in audio-based context recognition.Cepstrum coefficients extraction method based on the minimum variance distortionless response and perceptual information is studied. This method is an improved algorithm of Mel frequency cepstrum coefficients which can effectively improve the speaker recognition system in noisy conditions.A feature selection algorithm based on local discriminant basis is proposed. This method selects the feature subset using two discriminant criterions and evaluates the feature subset with the recognition accuracy. After gained the best size of feature subset in training scheme, feature subset of the audio samples can be online received in the recognition scheme. Experiments in audio-based context recognition show that the local discriminant basis based feature selection method can effectively improve the recognition accuracy.A discriminative training algorithm that uses support vector machines to improve the classification accuracy of the continuous hidden markov model (HMM) is studied. The algorithm trains HMM models as a baseline system and uses support vector machines training scheme to rescore the parameters of HMM. Experiments show that support vector machine rescoring of hidden markov models typically improve the recognition accuracy compared to the original HMM.