OpenAI has developed a voice imitation tool that can mimic human voices using just a 15-second audio sample. However, due to potential security risks, the tool has not been released yet.
On March 29, OpenAI, which also developed the AI platform ChatGPT, conducted a small-scale experiment with the Voice Engine tool, a week after trademarking it. This tool is capable of replicating a person’s speech from a 15-second audio clip.
OpenAI has plans to test the tool but has not made it widely available due to the potential for misuse. The company acknowledges that the ability to create speech that can sound like anyone carries significant risks.
OpenAI is working with partners in the U.S. and internationally across various sectors, including government, media, entertainment, and education, to gather feedback during the development of this voice tool.
While some startups have already released voice imitation technologies, there are concerns among researchers about the potential misuse of these technologies to spread misinformation. This is particularly concerning as voice replication tools become more common, affordable, easy to use, and harder to trace.
In response to these concerns, OpenAI has stated that they are proceeding with caution and thoughtfulness before releasing the Voice Engine tool. The company has set rules for testing partners, including the need for explicit and full consent from all individuals whose voices are mimicked. In addition, listeners are informed when the voice they hear is AI-generated.
OpenAI has also put in place a series of safety measures, such as watermarking to trace the origin of any sound generated by the Voice Engine, and active monitoring of its usage.