Speech recognition method and device for reducing command word misrecognition, equipment and medium

By training the CTC speech recognition model and filtering candidate paths through path search, calculating the CTC loss, and weighting and optimizing the model parameters, the problem of misrecognition of short command words in embedded devices is solved, and the recognition accuracy is improved.

CN120808762AActive Publication Date: 2025-10-17深圳市友杰智新科技有限公司
View PDF 5 Cites 1 Cited by

Patent Information

Application Number
CN202511304429.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-09-12
Publication Date
2025-10-17
Estimated Expiration
2045-09-12

AI Technical Summary

Technical Problem

When embedded devices use the CTC algorithm to recognize short command words and colloquial command words, the probability of misrecognition is high, which cannot meet users' requirements for recognition accuracy.

Method used

The initial CTC speech recognition model is trained until it converges. A path search algorithm is used to filter candidate paths. The negative of the CTC loss is calculated as the reference path score. The candidate path scores are compared with the reference path scores. Finally, the CTC loss and the contrast loss are weighted and summed. The model parameters are iteratively adjusted to improve the model's ability to distinguish short/colloquial command words.

Benefits of technology

It reduces the probability of command word misidentification, improves the accuracy of command word recognition in embedded device scenarios, and meets users' needs for recognition accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120808762A_ABST
    Figure CN120808762A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of voice recognition, and discloses a voice recognition method and device for reducing command word misrecognition, equipment and a medium. The method comprises the following steps: training an initial CTC speech recognition model until a verification set is converged to obtain a convergence model; screening a plurality of candidate paths through a path search algorithm based on an original output value output by the convergence model; then calculating the CTC loss of the reference path, taking an opposite number as a reference path score, and carrying out comparison operation on each candidate path score and the reference path score to obtain a comparison loss; and finally, performing weighted summation on the CTC loss and the comparison loss to obtain total loss, and iteratively adjusting convergence model parameters by using the total loss to obtain a target speech recognition model. According to the method, the discrimination degree of correct and wrong paths can be enhanced, the model discrimination capability is improved, the misrecognition probability is reduced, the requirement of a user for recognition accuracy is met, and the method is adaptive to an embedded equipment scene.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Acoustic model training method and device, computer equipment and storage medium

    CN111128137A

  • Model determination method and related device

    CN117828353A

  • Voice transcription method and device based on comparative learning, equipment and medium

    CN118471227A

  • Command word confirmation method and device, equipment and storage medium

    CN119811375A

  • Speech recognition method and system for constructing small language based on whispertoken

    CN120340494A

Cited By

  • Speech recognition method and device and terminal equipment

    CN121838738A