Improving Continuous Sign Language Recognition: Speech Recognition Techniques and System Design

摘要

Automatic sign language recognition (ASLR) is a special case of automatic speech recognition (ASR) and computer vision (CV) and is currently evolving from using artificial lab-generated data to using 'real-life' data. Although ASLR still struggles with feature extraction, it can benefit from techniques developed for ASR. We present a large-vocabulary ASLR system that is able to recognize sentences in continuous sign language and uses features extracted from standard single-view video cameras without using additional equipment. ASR techniques such as the multi-layer-perceptron (MLP) tandem approach, speaker adaptation, pronunciation modelling, and parallel hidden Markov models are investigated. We evaluate the influence of each system component on the recognition performance. On two publicly available large vocabulary. databases representing lab-data (25 signer, 455 sign vocabulary, 19k sentence) and unconstrained 'real-life' sign language (1 signer, 266 sign vocabulary, 351 sentences) we can achieve 22.1% respectively 38.6% WER.