A brand new AI voice-cloning instrument from a British agency claims to have the ability to reproduce a spread of UK accents extra precisely than a few of its US and Chinese language rivals.
As a result of a lot of the information historically used to coach AI merchandise with voices comes from North American or southern English talking sources, many synthetic voices are inclined to sound related.
To fight this, the corporate Synthesia spent a 12 months compiling its personal database of UK voices with regional accents, via recording individuals in studios and gathering on-line materials.
It used these to coach a product known as Categorical-Voice, which might clone an actual individual’s voice or generate an artificial voice.
These can be utilized in content material akin to coaching movies, gross sales help and shows.
The corporate stated its clients wished extra correct regional representations.
“In the event you’re the CEO of an organization, or if you happen to’re only a common individual, when you have got your likeness, you need your accent to be preserved,” stated Synthesia Head of Analysis Youssef Alami Mejjati.
He added French-speaking clients had additionally commented that artificial French voices tended to sound French-Canadian fairly than originating from France.
“That is simply because the businesses constructing these fashions are usually North American corporations, they usually are inclined to have datasets which might be biased in direction of the demographics that they are in,” he stated.
The toughest accents to imitate are the least frequent, Mr Mejjati stated, as a result of there’s much less recorded materials obtainable to coach an AI mannequin.
There are additionally reviews that voice-prompted AI merchandise, akin to good audio system, usually tend to battle to grasp a spread of accents.
Final 12 months, inner paperwork from West Midlands Police revealed worries about whether voice recognition systems would understand Brummie accents.
In the meantime the US-based start-up Sanas is taking the other method, creating instruments for deployment in name centres which “neutralise” the accents of Indian and Filipino workers, as reported by Bloomberg in March.
The agency says it goals to scale back “accent discrimination” skilled by employees when callers fail to grasp them.
There’s concern that languages and dialects are being misplaced within the digital period.
“Among the many over seven thousand languages that also exist right this moment, nearly half are endangered in line with UNESCO; a few third have some on-line presence; lower than 2 % are supported by Google Translate; and in line with OpenAI’s personal testing, solely fifteen, or 0.2 % are supported by GPT-4 [an OpenAI model] above an 80 % accuracy,” writes Karen Hao within the guide Empire of AI.
“Language fashions are homogenising speech,” agrees AI skilled Henry Ajder, who advises governments and tech corporations, together with Synthesia.
Nonetheless, the higher these merchandise develop into, the simpler they can even be within the palms of scammers.
Synthesia’s product won’t be free when it’s launched within the coming weeks, and may have guardrails round hate speech and express materials.
However there are already many free, open-source voice-cloning instruments that are simply accessible and fewer protected.
At the start of July, messages generated by an AI-cloned voice impersonating US Secretary of State Marco Rubio had been reported to have been despatched to ministers.
“The open supply panorama for voice has advanced so quickly over the past 9 to 12 months,” Mr Ajder provides.
“And that, from a security perspective, is an actual concern.”
