Vowel recognition for speaker independent Chinese speech recognition

Abstract

Vowel recognition is essential in Chinese speech recognition, especially in the speaker independent tasks. In this paper, the authors argued that the fixed length frame segmentation of the speech signal makes the feature extracting process lose essential features and introduces some irrelevant information so that the extracted features may be less expressive and consistent. Using the pitch-based dynamically adaptive frames will improve the process of extracting the speech features so that they can be more expressive for the phonemes to be recognized, and more consistent among different speakers. The algorithm for dynamically segmenting the speech signals is discussed, and a variety of features has been tested with the pitch-based adaptive frames, and a new type of feature, the FFT magnitude pattern, shows that it is very expressive and consistent and may help to simplify the recognition models. By the use of the FFT magnitude patterns, definite algorithms can be adopted at the recognition stage. This will simplify the calculation and speed up the process. The experiment is done using a finite-state machine model. The results showed that the pitch-based FFT magnitude patterns are more expressive and consistent than other features and suitable for speaker independent Chinese speech recognition tasks.

This work was supported by the Trans-Century Training Programme Foundation for the Talents by the State Education Commission, China.