Phonetic Pitch Detection And Tone Recognition Of The Continuous Chinese Three-syllabic Words

Along with the development of modern scientific and computer technology, more natural and convenient methods are pursued in the communication between human and the machines. Generally speaking, there are approximately two aspects in the language communication between human and the machines — speech composition and speech recognition. Speech recognition, which focuses on speech as object, is an important research direction of speech signal processing and a primary technology to realize the communication between human and machines. Speech recognition is applied to such fields as computer, information processing, communication and electric system and automatic control as well as industry, military affairs, traffic, medicine and civil usage. But because of the feature that Chinese holds, speech information processing in Chinese is more difficult and complex than some west languages. Just these reasons lead to many puzzles so that speech recognition could far from meet the practical application. So now something that needs to be solved is to seek a new speech recognition algorithm and improve the veracity of recognition.Chinese is a language with syllabic tone. Initial and compound vowel of a Chinese syllable and tone are the dominating attributes. Tone as one of them takes important discriminate acceptation information, which is a powerful means in the continuous syncopation. Moreover, it is also imperative in the tone composition and pattern of continuous language stream as well as improving the recognition efficiency of single word or sentence and speech comprehension. Tone has an important meaning not only in exploring new speech recognition methods that possess Chinese characters but also in the speaker recognition, which bears personal speech tone characters. So the research of tone characters in Chinese has universal meaning. The tone character of three-syllabic words is more approximate to the character of continuous syllable words, so in the dissertation an effective method of phonetic pitch detection and tone recognition for the continuous Chinese three-syllabic words is discussed. Veracity of the phonetic pitch detection has a great influence on tone recognition. With the central two parts as the research contents: phonetic pitch detection and tone recognition, frequency feature of the phonetic pitch is discussed and an effective algorithm of phonetic pitch detection is given. Moreover, a new tone recognition algorithm is also put forward on the basis of the existing tone recognition methods.First, the development of the speech recognition technology, the system frame of the speech recognition and the present difficulties of the speech recognition technology as well as the present situation and theory basis of the tone recognition are comprehensive summarized. Then the basic algorithm and the new algorithm are the major subjects in this dissertation:The first key to the technology: phonetic pitch detection. Phonetic pitch detection is a pivotal factor for the veracity of the tone recognition. The major solved problem is to accurately phonetic pitch frequency detection, especially the frequency of the continuous Chinese three-syllabic sonant words. A method of syllable segmentation is given, which the short time average energy and the short time over-zero rate are applied to effectively syllable segmentation and combined with autocorrelation function and average magnitude difference methods. Experiment is done to verify the improving efficiency of the method. The second key to the technology: tone recognition. On the basis of the result of phonetic pitch detection, an approach of fuzzy decision tree for tone recognition is discussed and a dynamic time alignment algorithm and improved neural network algorithm are presented. For different Chinese words and expressions or the same Chinese words and expressions that are spoken by diverse persons, frame numbers of the input Chinese words and expressions signals are different. But the input architectures of most of the neural network clas...