Pitch-robust acoustic feature using single frequency filtering for children’s KWS
Biswaranjan Pattanayak, Gayadhar Pradhan
Abstract
The pitch and speaking rate are the two significant factors that cause the acoustic mismatch in children’s keyword spotting (KWS) system. This paper proposes a pitch-robust acoustic feature based on single frequency filtering (SFF) for the development of children’s KWS system. In the proposed approach using SFF, the amplitude envelopes (AEs) of the speech data are computed at D-number of selected frequencies separated in Mel scale. The AEs are then averaged over short-time overlapping analysis frames and logarithmically compressed to represent the D-dimensional feature set per analysis frame, here termed as Mel spaced single frequency average log envelope (MSSF-ALE). By using the proposed MSSF-ALE feature, improved performance is observed for the deep neural network-hidden Markov model-based KWS system over the standard Mel-frequency cepstral coefficients (MFCC) and MFCC extracted from the smoothed spectra. The relative improvement of 104.44% in term-weighted value (TWV) for children’s KWS is observed over the MFCC by using MSSF-ALE. The performance of the KWS system is then evaluated with data-augmented training through explicit speaking rate modification of the training data set. The MSSF-ALE provides a relative improvement of 195.94% in TWV over MFCC with the data-augmented training. The MSSF-ALE also results in improved performance than the explored features in noisy test cases.