New Ai Voice Tool Trained To Copy British Regional Accents

Sedang Trending 6 jam yang lalu

Zoe Kleinman

Technology editor@zsk

A caller AI voice-cloning instrumentality from a British patient claims to beryllium capable to reproduce a scope of UK accents much accurately than immoderate of its US and Chinese rivals.

Because overmuch of nan information traditionally utilized to train AI products pinch voices comes from North American aliases confederate English speaking sources, galore artificial voices thin to sound similar.

To combat this, nan institution Synthesia spent a twelvemonth compiling its ain database of UK voices pinch location accents, done signaling group successful studios and gathering online material.

It utilized those to train a merchandise called Express-Voice, which tin clone a existent person's sound aliases make a synthetic voice.

These tin beryllium utilized successful contented specified arsenic training videos, income support and presentations.

The institution said its customers wanted much meticulous location representations.

"If you're nan CEO of a company, aliases if you're conscionable a regular person, erstwhile you person your likeness, you want your accent to beryllium preserved," said Synthesia Head of Research Youssef Alami Mejjati.

He added French-speaking customers had besides commented that synthetic French voices tended to sound French-Canadian alternatively than originating from France.

"This is conscionable because nan companies building these models thin to beryllium North American companies, and they thin to person datasets that are biased towards nan demographics that they're in," he said.

The hardest accents to mimic are nan slightest common, Mr Mejjati said, because location is little recorded worldly disposable to train an AI model.

There are besides reports that voice-prompted AI products, specified arsenic smart speakers, are much apt to struggle to understand a scope of accents.

Last year, soul documents from West Midlands Police revealed worries astir whether sound nickname systems would understand Brummie accents.

Meanwhile nan US-based start-up Sanas is taking nan other approach, processing devices for deployment successful telephone centres which "neutralise" nan accents of Indian and Filipino staff, as reported by Bloomberg successful March.

The patient says it intends to trim "accent discrimination" knowledgeable by workers erstwhile callers neglect to understand them.

There is interest that languages and dialects are being mislaid successful nan integer era.

"Among nan complete 7 1000 languages that still beryllium today, almost half are endangered according to UNESCO; astir a 3rd person immoderate online presence; little than 2 percent are supported by Google Translate; and according to OpenAI's ain testing, only fifteen, aliases 0.2 percent are supported by GPT-4 [an OpenAI model] supra an 80 percent accuracy," writes Karen Hao successful nan book Empire of AI.

"Language models are homogenising speech," agrees AI master Henry Ajder, who advises governments and tech firms, including Synthesia.

However, nan amended these products become, nan much effective they will besides beryllium successful nan hands of scammers.

Synthesia's merchandise will not beryllium free erstwhile it is released successful nan coming weeks, and will person guardrails astir dislike reside and definitive material.

But location are already galore free, open-source voice-cloning devices which are easy accessible and little protected.

At nan opening of July, messages generated by an AI-cloned sound impersonating US Secretary of State Marco Rubio were reported to person been sent to ministers.

"The unfastened root scenery for sound has evolved truthful quickly complete nan past 9 to 12 months," Mr Ajder adds.

"And that, from a information perspective, is simply a existent concern."

Selengkapnya