Wake-up audio determination method, device, equipment and storage medium

Abstract

The application discloses a wake-up audio determining method, device, equipment and storage medium, belonging to the technical field of speech. In the embodiment of the present application, wake-up audio and non-wake-up audio are respectively modeled, each corresponding to a plurality of sentence states, forming a sentence state sequence, so that when the audio features of the audio are classified, it can be determined that the audio is more like a wake-up Audio is still more like non-wake-up audio. In this process, the wake-up audio and non-wake-up audio are directly modeled, and the two are independent of each other, instead of modeling for each phoneme, so there is no need for a model trained for each frame-level annotation data, and the recognition process There is also no need to determine the corresponding recognition result for each phoneme, which can greatly reduce the amount of calculation and improve the recognition efficiency.

Description

technical field [0001] The present application relates to the field of voice technology, and in particular to a wake-up audio determination method, device, device and storage medium. Background technique [0002] In recent years, with the continuous development of audio processing technology, intelligent voice interaction systems such as smart speakers and vehicle-mounted voice interaction systems have become popular. In order to reduce user operations, a voice wake-up function is provided. By recognizing the collected voice, it is determined Whether it is a wake-up voice, and then the voice wake-up of the device can be realized. [0003] In the related art, the method for determining the wake-up speech is usually: performing feature extraction on the speech to be processed to obtain a fixed-length speech feature, and inputting it into the wake-up acoustic model for classification. The sample data required for the training of the wake-up acoustic model needs to have frame-l...

AI Technical Summary

Problems solved by technology

[0004] In the above method of obtaining labeled data through the trained acoustic model, the alignment result will greatly affect the performance of subsequent models

For example, if the performance of the alignment model is poor and the accuracy of the alignment result is low, the performance of the trained model will be poor if the alignment result with low accuracy is used as label data.

If you want to get labeled data with high accuracy, you need to use large-scale sample data to retrain the acoustic model, which is costly and inefficient

Images