Encoder-decoder models for recognition of Russian speech
Keywords:
speech recognition, neural networks, end-to-end models, machine learning, attention mechanism, encoder-decoder modelsAbstract
Problem: Classical systems of automatic speech recognition are traditionally built using an acoustic model based on hidden Markov
models and a statistical language model. Such systems demonstrate high recognition accuracy, but consist of several independent
complex parts, which can cause problems when building models. Recently, an end-to-end recognition method has been spread, using
deep artificial neural networks. This approach makes it easy to implement models using just one neural network. End-to-end models
often demonstrate better performance in terms of speed and accuracy of speech recognition. Purpose: Implementation of end-toend
models for the recognition of continuous Russian speech, their adjustment and comparison with hybrid base models in terms of
recognition accuracy and computational characteristics, such as the speed of learning and decoding. Methods: Creating an encoderdecoder
model of speech recognition using an attention mechanism; applying techniques of stabilization and regularization of neural
networks; augmentation of data for training; using parts of words as an output of a neural network. Results: An encoder-decoder
model was obtained using an attention mechanism for recognizing continuous Russian speech without extracting features or using
a language model. As elements of the output sequence, we used parts of words from the training set. The resulting model could not
surpass the basic hybrid models, but surpassed the other baseline end-to-end models, both in recognition accuracy and in decoding/
learning speed. The word recognition error was 24.17% and the decoding speed was 0.3 of the real time, which is 6% faster than the
baseline end-to-end model and 46% faster than the basic hybrid model. We showed that end-to-end models could work without language
models for the Russian language, while demonstrating a higher decoding speed than hybrid models. The resulting model was trained on
raw data without extracting any features. We found that for the Russian language the hybrid type of an attention mechanism gives the
best result compared to location-based or context-based attention mechanisms. Practical relevance: The resulting models require less
memory and less speech decoding time than the traditional hybrid models. That fact can allow them to be used locally on mobile devices
without using calculations on remote servers.
