Data-Adaptive Single-Pole Filtering of Magnitude Spectra for Robust Keyword Spotting

This paper proposes a simple and effective data-adaptive smoothing approach to suppress the pitch and environment-induced mismatches in keyword spotting (KWS) systems. In the proposed method, the magnitude spectra are smoothed by processing through a data-adaptive single-pole filter (DA-SPF) before computation of Mel frequency cepstral coefficients (MFCCs) to filter out the high-frequency components, which are mainly due to the pitch periodicity. The pole magnitude, which controls spectral smoothing, is changed adaptively for each analysis frame depending on the normalized spectral magnitude in 0–2500 Hz frequency band. The formant magnitude of the voiced sound units is predominant in this frequency band. Consequently, the magnitude spectra of pitch-sensitive voiced frames are relatively more smoothed than the non-voiced frames. When the KWS systems are developed using MFCCs extracted from the DA-SPF smoothed spectra, referred to as single-pole smoothed (SPS)-MFCCs, significantly improved KWS performances are observed in pitch and noise mismatched test conditions. The SPS-MFCCs result in a relative improvement of 86.12% on the DNN-HMM-based KWS system over the MFCCs baseline for pitch mismatched test conditions.