Adaptation of Unsupervised Term Discovery for Speech to Sign Languages
Korhan Polat, Murat Saraҫlar · 2020
Unsupervised spoken term discovery (UTD), which is an active research area in speech processing, aims to find repeating units in speech signal using only the signal itself. It is useful for finding terms in low resource languages. Sign languages can also be considered as low resource since annotated corpora are not abundant, thus arises the need for unsupervised methods. In this work, we adapt a well known UTD algorithm to work with sign language videos. Instead of speech features, we used visual features obtained from a pre-trained convolutional neural net. We used an annotated sign dataset in order to evaluate the performance of the algorithm. According to the metrics that are used in UTD for speech, we show that the algorithm for sign language videos achieves similar performance compared to UTD for speech.