Streaming coding and speech recognition method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention provides a streaming coding and speech recognition method and device, electronic equipment and a storage medium. The streaming coding method comprises the following steps: determining a third transformation sequence of a historical reference frame sequence based on a second transformation sequence of a previous fused frame sequence; performing splicing processing based on the first transformation sequence and the third transformation sequence to obtain a fourth transformation sequence of the current fusion frame sequence; determining at least one of a source sequence and a context sequence for an attention mechanism based on the fourth transformed sequence; and performing streaming coding on the current frame sequence based on the source sequence and the context sequence. The third transformation sequence of the historical reference frame sequence is determined based on the second transformation sequence of the previous fusion frame sequence, so that the data processing amount required by the change processing of the historical reference frame sequence in the linear transformation processing process is reduced, the data processing efficiency is improved, and the computing power requirement of equipment is reduced.

Description

technical field [0001] Embodiments of the present invention relate to the field of computer technologies, and in particular, to a method, apparatus, electronic device, and storage medium for streaming encoding and speech recognition. Background technique [0002] Automatic Speech Recognition (ASR) systems are widely used in various interface applications, such as voice command-based search or knowledge question answering applications. [0003] Sequence to Sequence based neural network models have gained widespread popularity in ASR techniques. The input of an end-to-end ASR system is usually a speech sequence, and the output is usually a text sequence. Compared with traditional ASR systems, it can simplify the system structure and avoid the linguistic expert knowledge required to build ASR systems. This end-to-end ASR system can directly learn the various parts and links of the speech recognizer. [0004] Sequence-to-sequence models for end-to-end ASR systems are mainly ba...

AI Technical Summary

Problems solved by technology

[0006] However, applying speech recognition models such as transformers to devices with poor computing power still has low computational efficiency

Images