AN ATTENTION BASED DEEP NEURAL NETWORK FOR AUTOMATIC LEXICAL STRESS DETECTION

Lexical stress detection is one of important tasks in self-directed language learning application. We address this task by leveraging two successful attention techniques in natural language processing, inner attention and self-attention. First, combined with LSTM to model time-series features, inner attention could extract most important information and then convert length-varying input into a fixed-length feature vector; Second, self-attention intrinsically supports words with different number of syllables as input to model contexture information. Besides, our model is straightforward to expand to include hand-crafted features to further improve performance, and also can be applied to similar tasks, such as pitch accent detector. Experiments on LibriSpeech, TedLium and a third self-recored datasets show the high performance of our proposed attention based neural network.