Improving the Performance of Zero-Resource Children’s ASR System through Formant and Duration Modification based Data Augmentation

Abstract

Developing an automatic speech recognition (ASR) system for children’s speech is extremely challenging due to the unavailability of data from the child domain for the majority of the languages. Consequently, in such zero-resource scenarios, we are forced to develop an ASR system using adults’ speech for transcribing data from child speakers. However, differences in formant frequencies and speaking-rate between the two groups of speakers degrade recognition performance. To reduce the said mismatch, out-of-domain data augmentation approaches based on formant and duration modification are proposed in this work. For that purpose, formant frequencies of adults’ speech training data are up-scaled using warping of linear predictive coding coefficients. Next, the speaking-rate of adults’ data is also increased through time-scale modification. Due to simultaneous altering of formant frequencies and duration of adults’ speech and then pooling the modified data into training, the acoustic mismatch due to the aforementioned factors gets reduced. This, in turn, enhances the recognition performance significantly. Additional improvement is obtained by combining the recently reported voice-conversion-based data augmentation technique with the proposed ones. On combining the proposed and voice-conversion-based data augmentation techniques, a relative reduction of nearly 32.3% in word error rate over the baseline is obtained.