This paper describes the keyword search system developed by the STC team in the framework of OpenKWS 2016 evaluation. The acoustic modeling techniques included i-vectors based speaker adaptation, multilingual speaker-dependent bottleneck features, and a combination of feedforward and recurrent neural networks. To improve the language model, we augmented the training data provided by the organizers with texts generated by the character-level recurrent neural networks trained on different data sets. This led to substantial reductions in the out-of-vocabulary (OOV) and word error rates. The OOV search problem was solved with the help of a novel approach based on lattice generated phone posteriors and a highly optimized decoder. This approach outperformed familiar OOV search implementations in terms of speed and demonstrated comparable or better search quality.

The system was among the top three systems in the evaluation.