Efficient Keyword Spotting Using Time Delay Neural Networks

Abstract

This paper describes a novel method of live keyword spotting using a two-stage time delay neural network. The model is trained using transfer learning: initial training with phone targets from a large speech corpus is followed by training with keyword targets from a smaller data set. The accuracy of the system is evaluated on two separate tasks. The first is the freely available Google Speech Commands dataset. The second is an in-house task specifically developed for keyword spotting. The results show significant improvements in false accept and false reject rates in both clean and noisy environments when compared with previously known techniques. Furthermore, we investigate various techniques to reduce computation in terms of multiplications per second of audio. Compared to recently published work, the proposed system provides up to 89% savings on computational complexity.

Voice Activation Systems for Embedded Devices: Systematic Literature Review

“…The majority of cited works use MFCC or Log Mel-filterbank features. In this area, we see the reduction of the inductive bias over the time: more and more recent papers like (Raziel and Hyun-Jin, 2018) or (Myer and Tomar, 2018) don not use DCT-step, probably because deep neural networks work reasonably well even with correlated features. We expect further simplification: using raw waveform or some unsupervised approach like contrastive predictive coding in Oord et al (2018).…”

“…Whole word (Morgan et al, 1990;Rose and Paul, 1990;Naylor et al, 1992;Rohlicek et al, 1993;Cuayáhuitl and Serridge, 2002;Baljekar et al, 2014;Chen et al, 2014a;Zehetner et al, 2014;Hou et al, 2016;Manor and Greenberg, 2017;Fernández-Marqués et al, 2018;Myer and Tomar, 2018) Monophone (Rose and Paul, 1990;Rohlicek et al, 1993;Cuayáhuitl and Serridge, 2002;Heracleous and Shimizu, 2003;Szöke et al, 2005;Lehtonen, 2005;Silaghi and Vargiya, 2005;Wöllmer et al, 2009b;Jansen and Niyogi, 2009a,c;Wöllmer et al, 2009a;Szöke et al, 2010;Shokri et al, 2011;Tabibian et al, 2011;Hou et al, 2016;Kumatani et al, 2017;Gruenstein et al, 2017;Tabibian et al, 2018;Myer and Tomar, 2018) Triphone (Rose and Paul, 1990;Szöke et al, 2005) Part of the word (Naylor et al, 1992;Li and Wang, 2014;Chen et al, 2014a) State unit (Zeppenfeld and Waibel, 1992) Part of the phoneme (Rohlicek et al, 1989;Kosonocky and Mammone, 1995;Leow et al, 2012) Syllable (Klemm et al, 1995;…”

“…Comparing to threshold (Morgan et al, 1990;Naylor et al, 1992;Junkawitsch et al, 1997;Keshet et al, 2009;Wöllmer et al, 2009b,a;Li and Wang, 2014;Chen et al, 2014a;Gruenstein et al, 2017;Benisty et al, 2018;Myer and Tomar, 2018) Viterby (Rose and Paul, 1990;Feng and Mazor, 1992;Wilcox and Bush, 1992;Rohlicek et al, 1993;Knill and Young, 1996;Junkawitsch et al, 1997;Zheng et al, 1999;Liu et al, 2000;Vasilache and Vasilache, 2009;Tabibian et al, 2011;Leow et al, 2012;Zhu et al, 2013;Kumatani et al, 2017;Ge and Yan, 2017;Sun et al, 2017) Forward-Backward algorithm (Wilcox and Bush, 1992;Rohlicek et al, 1993) DTW (Zeppenfeld and Waibel, 1992;Kosonocky and Mammone, 1995;Kurniawati et al, 2012;Zehetner et al, 2014;Hou et al, 2016) Likelihood ratio (Jansen and Niyogi, 2009c;Szöke et al, 2010) Fuzzy logic (Manor and Greenberg, 2017) Table 10 The metrics used in studied sources.…”

A Time Delay Neural Network with Shared Weight Self-Attention for Small-Footprint Keyword Spotting

“…When the confident score exceeds a threshold, the keyword is detected. This approach shows the small footprint, low computational cost, low latency, high performance, and draws much attention recently [4,5,6]. However, previous work still use several hundred thousand parameters to achieve state-of-the-art performance.…”

“…However, because of the number of hidden layers and filters, their best model still has more than 200K parameters. A stacked time delay neural network (TDNN) based model with transfer learning was proposed [5]. However, the stacked network architecture makes the model size large.…”