A Rescoring Approach for Keyword Search Using Lattice Context Information

In this paper we present a rescoring approach for keyword search (KWS) based on neural networks (NN). This approach exploits only the lattice context in a detected time interval instead of its corresponding audio. The most informative arcs in lattice context are selected and represented as a matrix, where words on arcs are represented in an embedding space with respect to their pronunciations. Then convolutional neural networks (CNNs) are employed to capture distinctive features from this matrix. A rescoring model is trained to minimize term-weighted sigmoid cross entropy so as to match the evaluation metric. Experiments on single-word queries show that lattice context brings complementary gains over normalized posterior scores. Performance on both in-vocabulary (IV) and out-of-vocabulary (OOV) queries are improved by combining NN-based scores with standard posterior scores.