Vowel Non-Vowel Based Spectral Warping and Time Scale Modification for Improvement in Children’s ASR

Abstract

Acoustic differences between children's and adults' speech causes the degradation in the automatic speech recognition system performance when system trained on adults' speech and tested on children's speech. The key acoustic mismatch factors are formant, speaking rate, and pitch. In this paper, we proposed a linear prediction based spectral warping method by using the knowledge of vowel and non-vowel regions in speech signals to mitigate the formant frequencies differences between child and adult speakers. The proposed method gives 31% relative improvement over the baseline system. We have also investigated time scale modification using RTISILA and SOLAFS algorithms and found that our proposed method performs better. Combining the proposed method with RTISILA and SOLAFS results in a further error rate reduction. The final combined system gives 49% relative improvement compared to the baseline system.

Spectral Modification Based Data Augmentation For Improving End-to-End ASR For Children's Speech

“…Both of these methods use a uniform warping factor for all the formants. In related work, LPC-based spectral warping was applied selectively on vowel and non-vowel locations along with time scale modification in [23]. More recently, a novel fundamental frequency-based frequency warping technique was proposed which was shown to improve the performance of children ASR when added with other data augmentation techniques [24].…”

“…2 (a) shows the mean scaling factor (%) for the first 3 formants, averaged across all the vowels of children speakers, with respect to adult male formants. Also, existing spectrum warping methods [20,21,23] do not consider spectral variabilities due to an underdeveloped vocal tract, such as varying formant's bandwidth and formant's energy with respect to the adult spectrum. Generally, these methods are used to normalize the children's speech or acoustic features to minimize the mismatch between train and test datasets but not as an augmentation.…”