SPEECH RECOGNITION METHOD AND RELATED APPARATUS

WIPO Patent Application WO/2024/094001

A1

Disclosed in the embodiments of the present application are a speech recognition method and a related apparatus. When speech recognition is performed, speech recognition is not necessarily performed on the basis of all hidden layers in a speech recognition model; instead, an accuracy prediction model predicts corresponding speech recognition accuracy according to hidden layer information outputted by each hidden layer that executes calculation, and if the speech recognition accuracy corresponding to an i-th hidden layer has satisfied a result output condition, this indicates that information calculated by the hidden layer sufficiently fits the actual text corresponding to the speech, and a speech recognition result is directly output on the basis of the hidden layer information of the i-th hidden layer, and subsequent iterative calculation is not performed by means of subsequent hidden layers any more. Therefore, with regard to speech for which the recognition difficulty is relatively low, an accurate speech recognition result can be output by means of fewer hidden layers performing iterative calculation, and it is not necessary to wait for the iterative calculation of subsequent multiple hidden layers, such that the overall speech recognition efficiency is improved while the speech recognition accuracy is ensured.

JIANG YONGJUN (CN)

PCT/CN2023/128483

May 10, 2024

October 31, 2023

Click for automatic bibliography generation  

TENCENT TECH SHENZHEN CO LTD (CN)

G10L15/06

US20210020167A1	2021-01-21
CN112542161A	2021-03-23
US20190122099A1	2019-04-25
CN109817213A	2019-05-28
US20170270910A1	2017-09-21

SHENPAT INTELLECTUAL PROPERTY AGENCY (CN)

View/Download PDF      PDF Help

Previous Patent: POWER MODULE FOR MOTOR

Next Patent: FRUCTOSE-1,6-DIPHOSPHATE OPTICAL PROBE, AND PREPARATION METHOD THEREFOR AND USE THEREOF