Training method and using method of voice wake-up hybrid model and related equipment

Abstract

The invention provides a training method and using method of a voice wake-up hybrid model and related equipment; during model training, a voice wake-up hybrid model is obtained through hybrid learningtraining of a voice separation network, a feature transformation network and a wake-up word detection network. When the model is used, an audio to be recognized is input into the voice wake-up hybridmodel, and the wake-up probability is directly obtained. When the wake-up probability is greater than the threshold, the wake-up word identified from a to-be-identified audio is judged. In a model training process, a first loss function obtained through the voice separation network and a second loss function obtained through the wake-up word detection network are weighted to obtain a comprehensive loss function, and weight parameters of the voice separation network, the feature transformation network and the wake-up word detection network are obtained through back propagation and learning according to the comprehensive loss function. Meanwhile, the networks are unified into one framework, and a joint optimization method is used, so that the model can learn optimal separation and wake-up network parameters at the same time, and the wake-up rate is effectively improved.

Description

technical field [0001] The present application relates to the technical field of voice wake-up, and in particular to a training method, usage method and related equipment of a voice wake-up hybrid model. Background technique [0002] In the existing voice wake-up technology, there is a method based on the traditional GMM-HMM (Gaussian Mixture Model-Hidden Markov Model), which uses monophone or triphone as the hidden state of HMM (Hidden Markov Model), and obtains the sound sequence After the signal, the optimal state sequence is obtained by decoding to judge whether there are target keywords in the voice signal. This method is complex to implement, and requires detailed labeling information for samples. The cost of data acquisition is high, and the effect is not ideal. Word rejection is poor; other end-to-end methods based on DNN (deep neural network) have a simpler implementation process and better results than traditional methods, but the models are generally too large to ...

AI Technical Summary

Problems solved by technology

[0003] The main purpose of this application is to provide a training method, usage method and related equipment of a voice wake-up hybrid model, aiming to solve the drawbacks of the existing voice wake-up technology with high complexity and low wake-up rate

Images