Acoustic model training method and device, electronic equipment and storage medium

Abstract

The invention relates to an acoustic model training method and device, electronic equipment and a storage medium. The method comprises the steps that multiple voices and texts corresponding to the multiple voices are acquired; training the plurality of voices and the texts corresponding to the plurality of voices to obtain a basic model; determining a forward training set and a reverse training set corresponding to the target voice according to the basic model; and training the forward training set and the reverse training set to obtain an acoustic model. The target voice is searched in the multiple voices through the basic model, the voices similar to the target voice in pronunciation are searched in the multiple voices, the target voice forms the forward training set, the voices similar to the target voice in pronunciation form the reverse training set, and the forward training set and the reverse training set are trained to obtain the acoustic model. The acoustic model not only recognizes the target voice, but also can recognize the voice which is confused with the target voice, so that the probability of false wake-up is reduced.

Description

technical field [0001] The present application relates to the technical field of model training, in particular to an acoustic model training method, device, electronic equipment and storage medium. Background technique [0002] For voice wake-up or command word recognition on edge devices, limited by resources, the volume of the acoustic model is relatively small and the performance is relatively weak. The general solution is to first use a large amount of labeled speech to train a basic model, then record some recordings of target words, and then use these recordings to do adaptive training on the basic model, so that the model is more biased towards accurate recognition of target words. [0003] In order to optimize the recognition rate, it is generally necessary to find at least hundreds of people to record the voice of the target word, which makes the model optimization process unable to be copied quickly and automatically, which affects the speed of project implementati...

AI Technical Summary

Problems solved by technology

[0003] In order to optimize the recognition rate, it is generally necessary to find at least hundreds of people to record the voice of the target word, which makes the model optimization process unable to be copied quickly and automatically, which affects the speed of project implementation

Different products often use personalized wake-up words and command words, so the accumulated recordings cannot fully meet the needs of new products

Moreover, recording also requires capital investment, increasing costs

Images