- New tool can detect an AI voice over.
- Spoofed audio identification.
- French IRCAM organisation behind research.
As one of the greatest impressionist acts of the 21st century to date, artificial intelligence can often be found being put to use imitating the human voice. In appropriate contexts, AI-generated voice can be incredibly useful – for example, helping those with visual impairment access information that’s only presented visually, or adding a level of realism to spoken translations that make the originating message clearer to its audience.
There is, however, naturally a downside. The easy-availability of speech engines that can be taught inflection and accent from audio samples of a ‘real-life’ human speaker means that not only do voice actors find themselves in much less demand than a few years ago, but people whose voices are well-recognised can find audio clones of themselves saying just about anything. Human audiences simply can’t distinguish between the simulacrum and the real.
To help combat the trend in AI voice over spoofs and emulation, Ircam Amplify, the commercial branch of the French audio research organisation, IRCAM, has developed an AI speech detection tool that can spot audio fakes, with what it claims as a 98% accuracy rate.
With hundreds of voice models freely available on the market, and a similar number of tools that will apply the characteristics of a given recording onto the output of artificial speech algorithms, the problems associated with impersonation will indubitably grow. Whether used solo or accompanying AI-spoofed video, an impersonation’s effects can range from audience titillation and amusement, right through to having significant effects on the outcome of political elections. Imaginative criminals may also use voice spoofing to carry out crime.
AI voice over:Â Ã la francaise
Ircam Amplify has offered an AI-generated music detector since May of this year, and IRCAM’s acknowledged prowess in all things acoustic and computational has added the fake speech detector to the portfolio of its commercial cousin. Users can analyse up to 20 audio files per API call, and the available compute can “process hundreds of calls simultaneously'” according to a press release dated October 29, 2024.
To produce the testing routines, IRCAM trained its solution on common voice AI models; five commercial offerings and four open-source. Ircam Amplify states the code can ingest new gen-AI models “in under three weeks.”
IRCAM (Institut de recherche et coordination acoustique/musique) was established in 1977 and headed initially by Pierre Boulez, a well-known (in contemporary classical music circles) avant-garde composer with a reputation for experimentalism. A key part of the Institut‘s work since its inception has been electro-acoustic music, that is, the treatment of sound by electronics and computing. It is located next door to Paris’ Pompidou Centre, reflecting its role as a tenet of French cultural life.
While musical output from the Institut may be regarded by all but the avant-garde cognoscenti as the aural equivalent of kicking a steel can of cats down a cobbled street, the IRCAM-originated Max/MSP programming language is still in common use today.
Drawing on the skills of many decades of audio research and experimentation at IRCAM, the Ircam Amplify platform should pave the way for fast and efficient verification of audio released online.
Policing the veracity of media that appears on the internet is a new, and therefore grey area legally. To take one example from many thousands, the current Governor of California, Mike Huckabee, is one of the surprisingly-few ‘celebrities’ taking legal action against a content delivery platform (Meta, in this particular case) for distributing spoofed media based on his visual (and audio) likeness. A Doppelgovernor has been seen and heard promoting CBD sweets on Facebook, in contradiction to the conservative politician’s personal views about marijuana.
