Real-time subtitle translation and system implementation method for live broadcast scene

Abstract

The invention discloses an implementation method of a real-time subtitle translation and a system for a live scene. The method comprises the following steps of: S1, training a deep convolutional neural network by using a training data set; S2, performing Fourier transform on the input voice of each frame, and taking time and frequency as the two dimensions of the image, to obtain a spectral map ofthe input voice signal; S3, inputting the obtained spectral map into the trained deep convolutional neural network, to obtain text data corresponding to the voice signal; and S4, displaying the textdata on a screen in real-time. According to the implementation method of the real-time subtitle translation and the system for the live scene, the speech of each person can be translated in real-timeand displayed on the screen in time, so that the writing efficiency is effectively improved, and a more excellent user experience is brought to the audience.

Description

technical field [0001] The present invention relates to a real-time translation method based on computer technology, in particular to a real-time subtitle translation for live broadcast scenes and a system implementation method, which belongs to the field of deep learning technology. Background technique [0002] In recent years, the webcasting industry and its related technologies have developed extremely rapidly. Various livestreaming platforms have emerged in an endless stream, and the content of livestreaming has also become increasingly rich. Not only are there individual livestreaming sessions on various topics, but also conferences and online courses in the form of livestreaming. . It is precisely based on this industry development trend that nowadays, in various live broadcast platforms, the demand for real-time translation and display of the live broadcaster's words is becoming stronger and stronger. [0003] At present, there are also some technical solutions for ...

AI Technical Summary

Problems solved by technology

Specifically, although the two-way LSTM technology can model the long-term correlation of speech, it is difficult to apply to the real-time translation of the live broadcast platform because it requires a complete sequence for speech recognition and has a high delay.

CNN technology has been applied to speech recognition around 2012, but because it uses fixed-length frame splicing as input, it cannot see enough speech context information, and CNN is used as a feature extractor, the convolution used There are few layers and limited expression ability, so it is difficult to apply to real-time translation of live broadcast platforms

Images