Speech processing method, speech encoder, speech decoder and speech recognition system

Abstract

The embodiment of the invention provides a speech processing method, a speech encoder, a speech decoder and a speech recognition system. The method comprises the following steps: acquiring a to-be-processed speech signal; processing the speech signal through a first neural network and a second neural network, obtaining first feature information and second feature information corresponding to the speech signals, wherein the calculation efficiency of the first neural network is higher than that of the second neural network, and the accuracy of the second feature information output by the second neural network is higher than that of the first feature information output by the first neural network; and according to the first feature information and the second feature information, determining target feature information for representing semantics in the speech signal. According to the technical scheme provided by the embodiment of the invention, the two pieces of feature information are obtained through the two different neural networks, and the two pieces of feature information have complementarity in speech processing efficiency and quality, so that the accuracy and reliability of obtaining the target feature information are improved.

Description

technical field [0001] The invention relates to the technical field of data processing, in particular to a speech processing method, a speech encoder, a speech decoder and a speech recognition system. Background technique [0002] Speech recognition technology can convert the voice waveform spoken by people into text that can be recognized by machines. For speech recognition technology, the speech recognition rate is an important indicator for evaluating speech recognition performance. In 2017, Google proposed a Transformer model that can perform speech recognition. Specifically, the Transformer model can use a text-related self-attention mechanism to perform speech modeling on the long-term correlation of speech to obtain a speech recognition model, and then The speech recognition operation is realized through the established speech recognition model. [0003] However, when the Transformer model uses the text-related self-attention mechanism to model the long-term correlat...

AI Technical Summary

Problems solved by technology

[0003] However, when the Transformer model uses the text-related self-attention mechanism to model the long-term correlation of speech, due to the large number of text-related parameters, the complexity of building a speech recognition model is high, and it also increases The difficulty of optimizing the speech recognition model, which greatly affects the quality and efficiency of speech signal recognition

Images