Pronunciation Evaluation Using Short And Long-term Features

Pronunciation evaluation has important application value in the computer-assisted speech rehabilitation and language teaching fields.Currently, most pronunciation evaluations at home and abroad are based on speech recognition. However, the relevance between pronunciation evaluation which are based on speech recognition technology and subjective evaluation of experts is relatively low. Therefore, we use the method of non-speech recognition to extract short-term and long-term features of the test speech and the reference speech, and evaluate pronunciation quality by comparing the differences of their characteristics. The main content of this article is as follows:Firstly, we present a pronunciation evaluation using short and long-term features scheme. The short-term feature is the speech feature with a frame speech signal unit, such as formant, pitch, energy, phrase length, Bark spectrum, MFCC spectrum, MFSC spectrum, HBSCC spectra characteristic parameters and so on in this article. Long-term feature is the speech feature extracted based on the speech section from the short-term feature envelope. In this paper, we selected some long-term parameters to express long-term features, such as the number of peak and valley points, the mean of short-term features, the increased rate and reduction ratio of the waveform And then, we integrate short and long-term features to create speech models from the acoustic, rhythm and perception of three aspects, and compute similarity between the models of the reference speech and the test speech. At length, the model similarity is mapped into objective score through BP network.Secondly, we study pronunciation evaluation methods based on acoustic features, prosodic features and perceptual features respectively. Also, we analysis different roles of acoustic features, prosodic features and perceptual features in pronunciation evaluation by experiment.Thirdly, in the selection and analysis of prosodic features, we present a prosody model which fuses pitch frequency, energy, and phrase length of three features, and add it to pronunciation evaluation.Fourthly, we analysis the roles of Bark spectrum, MFCC spectrum, MFSC spectrum, HBSCC spectrum and so on in pronunciation evaluation in the section of the perception feature. According to the experimental results, we select MFSC spectrum which is the most relevant perception feature to build the perception model.Fifthly, we merge the similarity of the acoustic, rhythm and perception models between the reference speech and the test speech to compute objective score through BP network. In addition, we analysis the relevance of this objective and subjective score by doing many experiments and give the experimental results.Finally, we design a prototype test system according to the research work and content above. Experimental results show that the pronunciation evaluation method can significantly improve the correlation between the objective evaluation and subjective evaluation.