Filterbank Learning for Noise-Robust Small-Footprint Keyword Spotting
This work addresses energy-efficient and noise-robust keyword spotting for deployment on low-resource devices, representing an incremental improvement over existing methods.
The study tackled the problem of improving noise robustness and energy efficiency in keyword spotting (KWS) by replacing handcrafted speech features with learned filterbanks, especially when reducing the number of channels. Results showed that using 8-channel learned features instead of 40-channel log-Mel features led to only a 3.5% relative accuracy loss and a 6.3x reduction in energy consumption on a noisy dataset.
In the context of keyword spotting (KWS), the replacement of handcrafted speech features by learnable features has not yielded superior KWS performance. In this study, we demonstrate that filterbank learning outperforms handcrafted speech features for KWS whenever the number of filterbank channels is severely decreased. Reducing the number of channels might yield certain KWS performance drop, but also a substantial energy consumption reduction, which is key when deploying common always-on KWS on low-resource devices. Experimental results on a noisy version of the Google Speech Commands Dataset show that filterbank learning adapts to noise characteristics to provide a higher degree of robustness to noise, especially when dropout is integrated. Thus, switching from typically used 40-channel log-Mel features to 8-channel learned features leads to a relative KWS accuracy loss of only 3.5% while simultaneously achieving a 6.3x energy consumption reduction.