Voice keyword recognition method and system

Abstract

The invention discloses a voice keyword recognition method and system. The method comprises the steps: obtaining a preset voice data set, extracting fbank features from the preset voice data set, carrying out the normalization processing of the fbank features, and obtaining the normalized features; and performing variable-length sequence model training on the normalized feature by adopting preset seq2seq + attention to obtain a voice recognition model, performing end-to-end model training adjustment on the voice recognition model by adopting a preset asr model to obtain a training model, and finally performing voice recognition through the training model. According to the invention, the open-source data set can be fully utilized for network pre-training, the problem that the correct recognition rate is reduced due to the fact that a large number of junk words are contained in the data set can be avoided, the problem that key words are difficult to adjust due to the fact that a framework of an original training model is too large can also be avoided, and therefore the recognition accuracy and efficiency are effectively improved; and the identification flexibility and practicability are improved.

Description

technical field [0001] The invention relates to the technical field of voice recognition, in particular to a voice keyword recognition method and system. Background technique [0002] With the rapid development of the Internet of Things and the increasing popularity of Internet of Things devices, in order to improve the interaction efficiency between users and users using Internet of Things devices, a speech recognition model is usually added to the device to recognize the user's intention and perform corresponding operations through the speech recognition model . [0003] The speech recognition method of the commonly used speech recognition model can be subdivided into ASR recognition with a large vocabulary, recognition with a small vocabulary and keyword recognition. Specifically, the training data set input by the user is used to perform neural network training on the above recognition method of the data set information, so as to generate a corresponding speech recognit...

AI Technical Summary

Problems solved by technology

[0004] However, the currently commonly used speech recognition models have the following technical problems: the data set required for recognition training is very demanding. If the number of keywords in the data set is small, the data set will contain a large number of irrelevant words, which will reduce the accuracy of recognition. The data set contains a large number of specific keywords for training, and the user needs to spend a lot of time collecting words in the early stage, which not only increases the difficulty of training, but also reduces the training efficiency; and because the trained speech recognition model is trained with a single data set, making speech recognition The model is difficult to adjust, further reducing the flexibility and practicability of the model

Images