基于深度学习的视频语音提取文本系统设计与实现
Design and Implementation of Video Speech Extraction Text System Based on Deep Learning

21世纪是信息化的时代，多媒体技术在网络教学中的应用越来越普及。在新冠疫情防控形势严峻的时期，网络教学凭借得天独厚的优势起到了重大作用。但当前市场上的在线视频编辑平台功能单一、效率低下、用户体验繁琐，本文利用基于循环神经网络和卷积神经网络实现的语音识别对视频进行文本提取，并且使用注意力机制算法实现的语音合成对视频文本的修改，使用FFmpeg对视频进行处理，同时使用多线程和异步队列提升系统性能。本文主要针对如何实现语音识别和语音合成，以及如何提升语音合成效果做了主要的研究，最终实现识别普通话的准确率为90.52%，以及声音合成近乎为ground truth的合成引擎，将其应用于产品实现了能对视频精准编辑，体验良好的视频语音提取文本系统。
The 21st century is an information age, and the application of multimedia technology in network teaching is becoming more and more popular. In the severe period of COVID-19 prevention and control, online teaching has played an important role by virtue of its unique advantages. However, the current online video editing platform in the market has the disadvantages of single function, low efficiency and cumbersome user experience. This paper uses speech recognition based on cy-clic neural network and convolutional neural network to extract the video text, uses speech syn-thesis realized by attention mechanism algorithm to modify the video text, and uses FFmpeg to process the video, it also uses multithreading and asynchronous queues to improve system per-formance. This paper mainly focuses on how to realize speech recognition and speech synthesis, and how to improve the effect of speech synthesis. Finally, the accuracy rate of Mandarin recogni-tion is 90.52%, and the sound synthesis engine is close to the ground truth, which can be applied to the product to realize accurate video editing and experience a good video-speech extraction text system.