F₀ Contour Analysis Based on Empirical Mode Decomposition for DNN Acoustic Modeling in Mandarin Speech Recognition

Xiaoyun Wang, Xugang Lu, Hisashi Kawai, Seiichi Yamamoto

Tone information provides a strong distinction for many ambiguous characters in Mandarin Chinese. The use of tonal acoustic units and F₀ related tonal features have been shown to be effective at improving the accuracy of Mandarin automatic speech recognition (ASR) systems, as F₀ contains the most prominent tonal information for distinguishing words that are phonemically identical. Both long-term temporal intonations and short-term quick variations coexist in F₀. Using untreated F₀ as an acoustic feature renders the F₀ contour patterns differently from their citation form and downplays the significance of tonal information in ASR. In this paper, we explore the empirical mode decomposition (EMD) on F₀ contours to reconstruct F₀ related tonal features with a view to removing the components that are irrelevant for Mandarin ASR.We investigate both GMM-HMM and DNN-HMM based acoustic modeling with the reconstructed tonal features. In comparison with the baseline systems using typical tonal features, our best system using reconstructed tonal features leads to a 4.5% relative word error rate reduction for the GMM-HMM system and a 3.5% relative word error rate reduction for the DNN-HMM system.

doi: 10.21437/Interspeech.2016-653

Cite as: Wang, X., Lu, X., Kawai, H., Yamamoto, S. (2016) F₀ Contour Analysis Based on Empirical Mode Decomposition for DNN Acoustic Modeling in Mandarin Speech Recognition. Proc. Interspeech 2016, 973-977, doi: 10.21437/Interspeech.2016-653

@inproceedings{wang16c_interspeech,
  author={Xiaoyun Wang and Xugang Lu and Hisashi Kawai and Seiichi Yamamoto},
  title={{F0 Contour Analysis Based on Empirical Mode Decomposition for DNN Acoustic Modeling in Mandarin Speech Recognition}},
  year=2016,
  booktitle={Proc. Interspeech 2016},
  pages={973--977},
  doi={10.21437/Interspeech.2016-653}
}