Exploring the role of pitch-adaptive cepstral features in context of children's mismatched ASR
Rohit Sinha, Syed Shahnawazuddin, Patri Satya Karthik · 2016
The presented work explores the role of pitch-adaptive cepstral features in context of automatic speech recognition (ASR) of children's speech on adults' speech trained acoustic models. On account of large acoustic mismatch between training and test data, highly degraded recognition rates are noted for such cases. Earlier studies have shown that the said acoustic mismatch is aided by the insufficient smoothing of pitch harmonics in the case of mel-frequency cepstral coefficient (MFCC) features for child speakers. Motivated by that, in this work, we explore pitch-adaptive cepstral features for reducing the sensitivity to gross pitch variations. For this purpose, a simple technique based on adaptive-cepstral-truncation is employed for deriving the pitch-adaptive MFCCs. We have also explored the existing STRAIGHT-based MFCCs for contrast. Both the approaches are found to result in significant and similar improvements for children's mismatch ASR case. The effectiveness of the adaptive-truncation-based approach is also demonstrated in context of the deep-neural-network-based acoustic models. Further, it has been shown that the effectiveness of the existing feature normalization techniques remain intact even with the use of the proposed features.