Analyzing pitch robustness of PMVDR and MFCC features for children's speech recognition
Shweta Ghai, Rohit Sinha · 2010
The degradation in children's speech recognition performance under mismatched condition i.e., on the adults' speech trained models is a well known problem. Apart from several other factors, this degradation is also contributed by the large difference in the pitch values of the adults' and the children's speech. MFCC is the most commonly used feature in automatic speech recognition but it has been reported to be affected by the pitch variations across speech signals. Recently, perceptual-MVDR (PMVDR) feature has been reported as a better alternative to MFCC under noisy conditions. It is also attributed to possess better spectral modeling ability for high pitch signals. Motivated by these, in this work, we analyze the robustness of PMVDR to pitch variations across speech signals in comparison to MFCC for the children's speech recognition under mismatched condition. Our study finds PMVDR to be more pitch robust than MFCC using the default parameters. However, on suitable adaptation of the parameters for the children's speech recognition under mismatched condition, both PMVDR and MFCC give significantly improved comparable performances for children's speech as well as exhibit similar robustness to pitch variations.