Sequence Summarizing Neural Networks for Spoken Language Recognition

This paper explores the use of Sequence Summarizing NeuralNetworks (SSNNs) as a variant of deep neural networks(DNNs) for classifying sequences. In this work, it is appliedto the task of spoken language recognition. Unlike other classificationtasks in speech processing where the DNN needs toproduce a per-frame output, language is considered constantduring an utterance. We introduce a summarization componentinto the DNN structure producing one set of language posteriorsper utterance. The training of the DNN is performed byan appropriately modified gradient-descent algorithm. In ourinitial experiments, the SSNN results are compared to a singlestate-of-the-art i-vector based baseline system with a similarcomplexity (i.e. no system fusion, etc.). For some conditions,SSNNs is able to provide performance comparable to the baselinesystem. Relative improvement up to 30% is obtained withthe score level fusion of the baseline and the SSNN systems.