HMM-Based Mandarin Tone Recognition - Application in Computer-Aided Language Learning System for Mandarin

Abstract:

The current paper reports our study on automatic Mandarin tone recognition towards the integration of tone recognition system in a computer-aided language learning (CALL) system for German learners of Mandarin. Three HMMbased tone recognition systems were developed including monotone, bitone and tritone recognizer for isolated monosyllabic, bisyllabic words and sentences, respectively. Different kinds of features, including prosodic and spectral-based features, were used in order to study its quality for tone recognition. The F0 contour was decomposed according to the Fujisaki model to its components which contain phrase components and tone components. In order to test the tone recognition systems on data from German learners of Mandarin, the tone models were adapted using correct data from the German students. The combination of prosodic and spectralbased features yielded better results than individual features. The results indicated that the proposed monotone and bitone recognizers outperform existing state-ofthe-art algorithms. The tone correctness of adapted acoustic models was better than original models.


Year: 2012
In session: Sprachtechnologie und Anwendungen
Pages: 347 to 354