Offline speech recognition learning method based on human behavior experience

Abstract

The invention discloses an offline speech recognition learning method based on human behavior experience, and the method comprises the steps: S01, segmenting a preset working time period into a plurality of interval time periods according to a preset interval time length, constructing an experience matrix corresponding to the plurality of interval time periods, and enabling the specification of the experience matrix to be M*N, assigning values to the data in the empirical matrix and associating corresponding parameters; S02, receiving a voice signal, importing the voice signal into a voice recognition system, outputting a voice recognition result, converting the voice recognition result into M*N matrix data, correspondingly calling experience matrix data of a corresponding interval periodaccording to a time node obtained by the voice signal, performing mathematical calculation on the experience matrix data, and outputting a calculation result; S03, when the calculation result meets apreset condition, judging that voice recognition succeeds, and otherwise, judging that voice recognition fails; and S04, adjusting and updating the empirical matrix data according to a voice recognition success or voice recognition failure result. According to the scheme, the adaptive ability is good, the response accuracy is high, and the false touch rate is low.

Description

technical field [0001] The invention relates to the technical field of speech recognition, in particular to an offline speech recognition learning method based on human behavior experience. Background technique [0002] The current mainstream speech recognition systems can be divided into online large-scale continuous vocabulary speech recognition and offline small-vocabulary speech recognition. The online mode relies on the powerful computing power of the server to not only support more scenes and more words, but also the recognition stability and accuracy. Can be well guaranteed. While offline small vocabulary speech is mostly deployed on embedded platforms with limited computing power, the size of the neural network model is limited, the expressive ability is also worse, and the performance of the resulting speech recognition system is limited. [0003] Among them, the operation mechanism of traditional small vocabulary speech recognition is briefly introduced as follows...

AI Technical Summary

Problems solved by technology

The expression ability of the compressed model is greatly reduced. Not only can it only do limited command word recognition on the embedded side, but also the generalization ability of the model is limited. Not only the recognition accuracy is affected, but also the false trigger is high.

[0007] In addition, in the current field of embedded terminal offline speech recognition system, because the limited model affects the performance of speech recognition, users often set different sensitivities to make the system conform to the actual use situation, but the confidence of speech recognition is easily affected by noise. Interference, too complicated by users and prone to bad user experience

[0008] Due to limited resources, the speech recognition system is also susceptible to noise interference. When there is very noisy environmental noise, sound recognition is prone to errors, and the accuracy rate will also decrease.

Images