Recently, OpenAI revealed that it had carried out a limited trial of a novel instrument named Voice Engine. This speech cloning system analyzes a 15-second audio sample to replicate any speaker. The firm claims to provide “emotive and realistic voices” and “natural-sounding speech.”
The technology has been under development since 2022 and is predicated on the company’s current text-to-speech API. A subset of the toolkit has already been utilized by OpenAI to power the read-aloud function and the preset voices found in the existing text-to-speech API. The company’s official blog has a number of samples that sound uncannily similar to the original. I urge you to listen to them and consider the potential outcomes, both positive and negative.
OpenAI says they see this technology being useful for reading assistance, language translation and helping those who suffer from sudden or degenerative speech conditions. The company brought up a Brown University pilot program that helped a patient with speech impairment issues by creating a Voice Engine clone pulled from audio recorded for a school project.
Despite the potential benefits, bad actors would certainly abuse this technology to engage in some serious deepfake tomfoolery, which is already a problem. With this in mind, Voice Engine isn’t quite ready for prime time, as there are serious privacy concerns that must be met before a full rollout.
OpenAI acknowledges that this tech has “serious risks, which are especially top of mind in an election year.” The company says its incorporating feedback from “US and international partners from across government, media, entertainment, education, civil society and beyond” to ensure the product launches with a minimal amount of risk. All preview testers agreed to OpenAI’s usage policies, which ban the impersonation of another individual without consent or legal right.
Additionally, anybody using the tech will have to disclose to their audience that the voices are AI-generated. OpenAI implemented safety measures, like watermarking to trace the origin of any audio and “proactive monitoring” of how the system is being used. When the product officially rolls out there will be a “no-go voice list” that detects and prevents AI-generated speakers that are too similar to prominent figures.
As for when that rollout will occur, OpenAI remains tight-lipped. TechCrunch uncovered some potential pricing data and it looks like it will undercut competitors in the space like ElevenLabs. Voice Engine could cost $15 per one million characters, which works out to around 162,500 words. This is about the length of Stephen King’s The Shining. It certainly sounds like a budget-friendly way to get an audiobook done. The marketing materials also make reference to an “HD” version that costs twice as much, but the company hasn’t detailed how that will work.
OpenAI has been making big moves this week. It just announced another partnership with its bestie Microsoft to build an AI-based supercomputer called “Stargate.” The project will reportedly cost a whopping $100 billion, according to The Information.
