Investigating Joint CTC-Attention Models for End-to-End Russian Speech Recognition

摘要

We propose an application of attention-based models for automatic recognition of continuous Russian speech. We experimented with three types of attention mechanism, data augmentation based on a tempo and pitch perturbations, and a beam search pruning method. Moreover we propose a using of sparsemax function for our task as a probability distribution generator for an attention mechanism. We experimented with a joint, CTC-Attention encoder-decoders using deep convo-lutional networks to compress input features or waveform spectrograms. Also we experimented with Highway LSTM model as an encoder. We performed experiments with a small dataset of Russian speech with total duration of more than 60 h. We got the recognition accuracy improvement by using proposed methods and showed better performance in terms of speech decoding speed using the beam search optimization method.