Voice recognition method and device

Abstract

The invention provides a voice recognition method. The method comprises the following steps: adding scene classification information of an acoustic model; obtaining acoustic model output of to-be-tested voice under the scene classification information, and determining a scene meeting a condition; and dynamically loading the voice model corresponding to the scene meeting the condition to obtain a voice recognition result. By applying the voice recognition method provided by the embodiment of the invention, scene information is added to the acoustic model, the scene model meeting the condition is dynamically loaded, the limitation of the original preset scene model is removed, the recognition efficiency is improved, and a dynamic loading mode is adopted after the model numerical value of therecognition scene is set. Therefore, the numerical value is not changed due to the change of the service demand, so that the response speed is stabilized at the decoding speed of the model with the set numerical value.

Description

technical field [0001] The invention relates to the technical field of data processing, in particular to a voice recognition method and device. Background technique [0002] Automatic Speech Recognition (ASR) consists of three parts: acoustic model, language model and decoder, as follows figure 1 shown. Among them, the acoustic model and the language model have their own training methods. The acoustic model uses speech data to train the sound-mapping pronunciation model; the language model uses text data to train the pronunciation-mapping text model. Generally, the language model will pre-train multiple models according to the usage scenarios. Use the scene to load the scene model that may be used; the two can be trained separately and in parallel; when using the ASR recognition project, it is currently necessary to manually set the boundary of the scene, that is, the acoustics need to be configured with a near-speaking scene or a far-speaking scene, and the language model ...

AI Technical Summary

Problems solved by technology

[0006] The purpose of the embodiments of the present invention is to provide a method and device for speech recognition to solve the problems of ASR recognition in the prior art that requires multiple language models to be preset, the recognition response speed is reduced, and the recognition scene does not match.

Images