With only 3.7 seconds of audio, a new AI algorithm developed by Chinese technology giant Baidu can clone a very reliable false sound. Just like the rapid development of machine learning software, this software can democratize the production of virtual video, and this research shows why it is increasingly difficult to believe in any media on the Internet.
The tech giant's researchers released their latest developments in Deep Voice, a system developed for sound cloning. A year ago, the technology required about 30 minutes of audio to create a new fake audio clip. Now, with just a few seconds of training material, it can create better results.
Baidu recently announced that Deep Voice, a new AI algorithm developed by Baidu, can perfectly clone a person's voice through 3.7 seconds of recorded sample data.
Deep Voice is a high-quality voice-transfer (TTS) system built by deep neural networks from Baidu AI Research Institute. The system not only improves the simulation time, but Baidu also optimizes the probability of its error. Even on a single GPU server, the inference scale was increased to more than 10 million times a day.
Application of adaptive speaker coding method in training, cloning and audio generation
Deep Voice was first released in the first edition in early 2017. The first version of the system can simulate the initial short sentences, and it is almost impossible to distinguish the difference from the real person. But the system can only simulate one person's voice at a time, and it takes several hours of learning to clone successfully. But the success of the latest release has been shortened to 3.7 seconds, and the female voice can be turned into a male, and the British voice becomes an American.
Simulator encoder structure
Researchers at Baidu Research Institute published the latest development of Deep Voice System "Neural Voice Cloning with a Few Samples" on the pre-printed website arxiv. In addition to cloning sounds with a small sample, the system can turn female voices into males and English sounds into Americans. Baidu researchers said the study could be applied to the personalization of human-computer interaction.
Metal Glasses,Metal Frame Glasses,Metal Eyeglasses,Retro Metal Frame Glasses
Danyang Hengshi Optical Glasses Co., Ltd. , https://www.hengshi-optical.com