Chinese Phonetic Tone Recognition Methods Based On Biomimetic Pattern Recognition

In Chinese, tone plays an important role in sense distinction. Syllables with the same consonant and vowel but with different tones may have a completely different meaning. For speech recognition, voice data need to be cut to extract necessary voice segment from each sub-paragraph speech. For that, this thesis combines short-time zero-crossing rate, the short-term energy method and frame cross-correlation method to cut the voice data. A series experiments were done afterwards. Experimental results show that this method can cut voiced consonants and vowels, low-noise and speech segments well,Pitch trajectory reflects the tone features about voice. Accuracy to extract pitch frequency influences the tone recognition. This thesis proposes a method to extract pitch frequency based on combination of the self-correlation method and time-domain features. Then a method combining three thresholds discard and replacement, median filtering and linear smoothing is applied for pitch track smoothing. As the sequence length varies, in order to unify pitch sequence length, this thesis also proposes a method by interpolation and structured uniform sequence length, which is facilitate for tone recognition later.This thesis applies biomimetic pattern recognition to Chinese tone recognition. A method to find coverage area and its regional centers by dynamically clustering and identify tone with multi-weight neuron coverage is proposed. At the same time experimental small samples and large samples were used to compare with HMM, SVM and BP algorithms. Experimental results show that the method proposed in this thesis has a good result.In the end, this thesis discusses about speech recognition modeling primitives and proposes the tree model based on consonants, vowels and tones. Then the model is applied to isolated word and words recognition. Good results are achieved in the last.