Model training method and device based on reversible separation convolution and computer equipment

Abstract

The invention relates to a model training method based on reversible separation convolution. The model comprises a network for processing audios and a network for processing texts; the network for processing the audio and the network for processing the text respectively comprise a reversible separation convolution layer; the method comprises the following steps: acquiring voice data of a specifieddata pair in a network computing training set for processing the audio, obtaining a first high-dimensional vector, obtaining text data of a network calculation specified data pair for processing a text, and obtaining a second high-dimensional vector, wherein the training set is composed of a data pair formed by voice data and text data, and the specified data pair is any data pair in the trainingset; training an audio processing network and a text processing network on the training set through a loss function, wherein the loss function is a spatial distance between the first high-dimensionalvector and the second high-dimensional vector; judging whether the loss function reaches a minimum value or not; and if so, judging that the training converges to obtain a twin network structure composed of the network for processing the audio and the network for processing the text. The model construction and training cost is saved.

Description

technical field [0001] This application relates to the field of computers, in particular to a model training method, device and computer equipment based on reversible separation convolution. Background technique [0002] "Speech recognition" and "speech synthesis" are two "sequence-to-sequence" prediction tasks in a dual relationship, which can be modeled using the encoder-decoder framework. Since the training data of "speech recognition" and "speech synthesis" are not universal, the existing speech recognition system only achieves a one-way mapping of aligning speech information to text information, and speech synthesis only achieves aligning text information to speech information one-way mapping. Due to the diversity of sequences, the scale of each one-way mapping system is very large, and the amount of data required for training the system is also very large, so the construction and training costs of each one-way mapping system are very high, which is not conducive to "s...

AI Technical Summary

Problems solved by technology

[0003] The main purpose of this application is to provide a model training method based on reversible separable convolution, which aims to solve the problem that the construction and training costs of the existing one-way mapping system are very high, which is not conducive to the general promotion of "speech recognition" and "speech synthesis" systems Technical issues used

Images