Design And Implementation Of The Mandarin Tone Recognition And Evaluation System

Tone is the most important information of the Mandarin. With the technology development of Chinese speech recognition, study on the tone recognition has been becoming a major way of technology breakthrough of Chinese speech recognition. The main characteristic of tone is fundamental frequency F0 and the effective approach to make use of tone information is to abstract and analyze F0 accurately. So the three aspects which include the pitch extraction and post-disposal, modeling of tone recognition and tone assessment has been designed in Mandarin system of tone recognition and assessment in this paper.Firstly, the paper has started from F0 abstraction and the extraction method is the short-term self-correlation method. Due to the inevitable oddity points during the tone extraction, the tone extraction post-disposal is important (necessary). The paper adopted the disposal of smoothly interpolation and resample to get the standard pitch information that can recognize models. Meanwhile, an innovation method (m_LD arithmetic) has been proposed, so that it can be more better to apply to the speaker-independent tone recognition. According to the speech database, the arithmetic experiment result has obtained and the validity of arithmetic has been proved.Secondly, on the basis of extracting the pitch information of which the paper has analyzed the characteristics and collectivity distribution, so it can get the fact that the collectivity distribution meets the Gaussian. The above work has provided a better theoretic foundation for tone recognition modeling. Thanks to the backup of the theory, we built the tone recognition model (GMM), analyzed the model in Standards speech database of 863, and presented the analyzing results.Thirdly, according to the tone recognition result of the GMM model, the paper builds Chinese tone assessment model which is based on the GMM Posterior probability and suggests some different kinds of score methods. The construction of assessment model is syncretized well with the tone pitch characteristics and recognition results. It has used two indexes of correlation and grade error rate, scaling the advantages and disadvantages of Chinese tone assessment. With the 100 university student's speech database and the nine language expert's scoring, it has presented the results of the corresponding indicators. Analysis of the results of experiment, the article has gived the following conclusion: In the system of tone recognition and tone assessment, the article has adopted the algorithm of Fundamental Frequency extract, post-disposal, tone recognition model and tone assessment model can achieve good effect.Finally, the paper has pointed out and concluded the shortage, and prospected the future.