Keyword spotting (KWS) aims to detect predefined keywords in continuous speech. Recently, direct deep learning approaches have been used for KWS and achieved great success. However, these approaches mostly assume fixed keyword vocabulary and require significant retraining efforts if new keywords are to be detected. For unrestricted vocabulary, HMM based keywordfiller framework is still the mainstream technique. In this paper, a novel deep learning approach is proposed for unrestricted vocabulary KWS based on Connectionist Temporal Classification (CTC) with Long Short-Term Memory (LSTM). Here, an LSTM is trained to discriminant phones with the CTC criterion. During KWS, an arbitrary keyword can be specified and it is represented by one or more phone sequences. Due to the property of peaky phone posteriors of CTC, the LSTM can produce a phone lattice. Then, a fast substring matching algorithm based on minimum edit distance is used to search the keyword phone sequence on the phone lattice. The approach is highly efficient and vocabulary independent. Experiments showed that the proposed approach can achieve significantly better results compared to a DNN-HMM based keyword-filler decoding system. In addition, the proposed approach is also more efficient than the DNN-HMM KWS baseline.
Multi-passage reading comprehension requires the ability to combine cross-passage information and reason over multiple passages to infer the answer. In this paper, we introduce the Dynamic Self-attention Network (Dyn-SAN) for multi-passage reading comprehension task, which processes cross-passage information at token-level and meanwhile avoids substantial computational costs. The core module of the dynamic self-attention is a proposed gated token selection mechanism, which dynamically selects important tokens from a sequence. These chosen tokens will attend to each other via a self-attention mechanism to model long-range dependencies. Besides, convolutional layers are combined with the dynamic self-attention to enhance the model's capacity of extracting local semantic. The experimental results show that the proposed DynSAN achieves new state-of-the-art performance on the SearchQA, Quasar-T and Wiki-Hop datasets. Further ablation study also validates the effectiveness of our model components.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.