CREU Blog 2017 - 2018 Research of JN Matthews

Keyword Dectection Literature Review

19 Nov 2017

Summaries of a few of the papers I’ve found on existing keyword spotting (KWS) are below:

  • Chen, Guoguo, Carolina Parada, and Georg Heigold. “Small-footprint keyword spotting using deep neural networks.” Acoustics, Speech and Signal Processing (ICASSP), 2014 IEEE International Conference on. IEEE, 2014.
    • They use a deep neural network to spot keywords instead of the traditional Hidden Markov Model. They found that it is a smaller implementation that has reduced computation time/space, and preforms at least as well.
    • Feature Extraction:
      • They use VAD as to only run the neural network on audio regions containing speech.
      • The audio features are captured in frames every 10 ms, and each encapsulate a 25 ms window.
      • The neutral net is given the 30 previous frames, and the future 10 frames at each instance.
    • Training
      • To train the network they had 2.3K instances of each keyword and 133K negative examples.
      • Keywords examined: answer call, ok google, take a picture, etc.
  • Hout, Julien van, et al. “Recent Improvements in SRI’s Keyword Detection System for Noisy Audio.” Fifteenth Annual Conference of the International Speech Communication Association. 2014.
    • They use a series of noise-robust features to improve the accuracy of their KWS in noisy conditions.
    • Noise Robust Features:
      • Damped Oscillator Coefficients
      • Normalized Modulation Coefficients
      • Modulation of Medium Duration Speech Amplitudes
      • Gammatone Filter Coefficients
      • Log-spectrally Enhanced Power Normalized Cepstral Coefficients
      • Gabor-MFCC
    • Results:
      • They found that the phoneme level search missing fewer keywords, than the word level search, but often has more false positives as a result.
      • They determined that a fusion of the features was most accurate for their tests.
  • Gales, Mark JF, et al. “Speech recognition and keyword spotting for low-resource languages: Babel project research at CUED.” SLTU. 2014.
    • They examine a zero resource acoustic model for languages with little to no transcribed audio.
    • They aim to build a language independent acoustic model, using features such as pitch, plp, and context dependent HMMs.
    • Their results were inconclusive and they found that although this idea is likely still possible there is more work to be done in the area.