Training method and device of wake-up model and computer equipment

Abstract

The invention relates to the field of artificial intelligence, and discloses a wake-up model training method. The method comprises the steps of extracting audio frames from specified voice statementsin a training set to obtain an acoustic feature matrix; inputting the acoustic feature matrix into a keyword detector of a first model to obtain a first spatial feature, inputting the acoustic featurematrix into an encoder of a second model to obtain a second spatial feature, the first model being a wake-up model to be trained, and the second model being a trained noise reduction model; calculating the difference between the spatial features of the first spatial feature and the second spatial feature; according to the calculation mode of the difference of the spatial features corresponding tothe specified voice statement, calculating the difference of the spatial features corresponding to all voice statements in the training set; and forming a loss function training wake-up model for training the wake-up model according to the differences of the spatial features corresponding to all the voice sentences and the preset cross entropy loss of the wake-up model. And the feature vector ofthe high-dimensional space is used as a knowledge distillation sample to assist in training a wake-up model, so that the wake-up effect is improved.

Description

technical field [0001] The present application relates to the field of artificial intelligence, in particular to the training method, device and computer equipment of the wake-up model. Background technique [0002] How to reduce the false wake-up rate has always been the main problem to be solved in the wake-up model. The general idea starts from two aspects. On the one hand, noise processing is performed in the data set. The noise data includes data of a specific scene or as many types of noise data as possible to simulate Real scene; Usually, the increase in the amount and type of noise data means that the network has a stronger learning ability, so when processing data, a more effective network structure should be designed on the model structure to improve the learning ability of the wake-up model. On the other hand, pre-processing modules are added before wake-up, including but not limited to traditional front-end gain amplification, reverberation, array noise reduction...

AI Technical Summary

Problems solved by technology

[0003] The main purpose of this application is to provide a training method for wake-up models, aiming to solve the technical problem that existing methods for reducing false wake-up rates cannot be applied to small-volume wake-up models

Images