OpenAI’s new text-to-speech technology can replicate voices

OpenAI’s text-to-speech technology, in contrast to earlier attempts at audio creation, can precisely reproduce individual voices using only 15 seconds of recorded audio. Now, it can also translate produced audio into multiple languages, which is advantageous for businesses such as Spotify. About ten developers are being shown early trials and previews of a new text-to-speech technology called Voice Engine by OpenAI. The company had initially planned to release the tool to as many as 100 developers through an application process.

"We recognize that generating speech that resembles people's voices has serious risks, which are especially top of mind in an election year. We are engaging with US and international partners from across government, media, entertainment, education, civil society and beyond to ensure we are incorporating their feedback as we build," the company wrote in a blog post.

In the testing program, OpenAI is requiring its partners to agree to its usage policies, obtain consent from the original speaker before using their voice, and to disclose to listeners that the voices they're hearing are AI-generated. The company is also installing an inaudible audio watermark to allow it to distinguish whether a piece of audio was created by its tool.

Before deciding whether to release the feature more broadly, OpenAI said it's soliciting feedback from outside experts. OpenAI also wrote that it hopes the preview of its software "motivates the need to bolster societal resilience" against the challenges brought about by more advanced AI technologies.   

One of OpenAI's current developer partners using the tool, the Norman Prince Neurosciences Institute at the not-for-profit health system Lifespan, is using technology to help patients recover their voice. For example, the tool was used to restore the voice of a young patient who lost her ability to speak clearly due to a brain tumor by replicating her speech from an earlier recording for a school project.