Learning Efficient Representations for Keyword Spotting with Triplet Loss

In the past few years, triplet loss-based metric embeddings have become a de-facto standard for several important computer vision problems, most no-tably, person reidentification. On other hand, in area of speech recognition generated by loss are rarely used even classification problems. We fill this gap showing that combination two representation learning techniques: embedding and variant kNN instead cross-entropy significantly (by 26% to 38%) improves accuracy convolutional networks on LibriSpeech-derived LibriWords datasets. To do so, we propose novel phonetic similarity based mining approach. also improve current best published SOTA Google Speech Commands dataset V1 10+2 -class about 34%, achieving 98.55% accuracy, V2 10+2-class 20%, 98.37% 35-class over 50%, 97.0% accuracy.