Exploration of MFCC based ABWE for robust children's speech recognition under mismatched condition

Abstract

In our previous works, it is demonstrated that the automatic speech recognition (ASR) performance for telephoneband children's speech in mismatched condition could be successfully improved following an existing LPC feature based artificial bandwidth extension (ABWE) method. Motivated by recent works reporting improved ABWE performance with MFCC features, in this work we explore the MFCC features for ABWE of children's speech and also present a novel algorithm for deriving the bandwidth extended MFCC features for ASR purpose without the conversion to speech domain. The proposed approach has much lower complexity and performs similarly when compared with the default speech domain approach. To address the higher variability in children's speech, we have also explored the age-specific conditioning and the effect of inclusion of memory (delta features) in ABWE modeling for children's speech.