Convolutional Recurrent Neural Networks for Small-Footprint Keyword Spotting

Abstract

Propose CRNN for KWS task (advancement of CNN for KWS)
With only ~230k parameters, CRNN model has low latency, achieves 97.71% accuracy at 0.5 FA/hour for 5 dB signal-to-noise ratio

Previous Works
- DNN with CE loss : not optimal performance
- CNN with CE loss : good performance, but no global view, can be improved
- RNN with CTC loss : no improvement
- CRNN with CTC loss by Lengerich et al. 2016 : no improvement
- So, we propose CRNN with CE loss which brings improvement!
End-to-End Architecture
Experiments
- developed own KWS system for the keyword “TalkType”
- ~16k samples with 5k different speakers
- frame length : 1.5 sec
- sampling rate : 16 kHz, each frame 24k time-domain sample
- PCEN (per channel energy normalized) mel spectrograms with 10 ms stride and 40 channels
- yields input dimensionality 40 x 151
- Data Augmentation
- additive noise, with a power determined by a signal-to-noise ratio (SNR) sampled from [-5,15] dB interval
- random timing jitter
- Evaluation
- false rejection rate (FRR) with fixed false alarms (FA) per hour 1 FA/hr
Choice of Model Architecture
- larger models typically perform better
- increasing conv filters or rnn hidden units are key factors
- GRU > LSTM
Noise Robustness
- Better generalization with higher SNR
Far-field Robustness
- voice input from distance generally performs worse due to train/test mismatch
- recovered by far-field augmentation