A groundbreaking AI voice-cloning technology from a British company promises to reproduce the full spectrum of UK regional accents with unprecedented accuracy, outperforming many of its American and Chinese competitors.
The challenge stems from traditional AI voice training data, which has historically drawn heavily from North American and southern English sources, resulting in artificial voices that sound remarkably homogeneous.
To address this limitation, London-based company Synthesia invested an entire year developing its own comprehensive database of UK voices representing diverse regional accents. The project involved extensive studio recordings and careful curation of online audio material from across Britain.
This rich dataset now powers Express-Voice, Synthesia’s latest innovation that can either replicate an individual’s authentic voice or generate entirely synthetic voices with regional authenticity.
The technology has broad applications across training videos, sales presentations, and corporate communications.
According to Synthesia, client demand drove this regional focus.
“Whether you’re a company CEO or an everyday individual, when your likeness is being represented, preserving your authentic accent becomes paramount,” explained Youssef Alami Mejjati, Synthesia’s Head of Research.
He noted that French-speaking clients had similarly complained that synthetic French voices consistently sounded French-Canadian rather than authentically European French.
“This bias exists simply because the companies developing these models are predominantly North American, working with datasets that naturally reflect their local demographics,” Mejjati observed.
The most challenging accents to replicate are invariably the rarest ones, Mejjati explained, as limited recorded material provides insufficient training data for AI models.
This accent recognition challenge extends beyond voice generation. Voice-activated AI products, including smart speakers, frequently struggle to comprehend diverse regional accents.
In a notable example from last year, internal West Midlands Police documents revealed concerns about whether voice recognition systems could effectively understand Birmingham accents.
Meanwhile, US startup Sanas has pursued the opposite strategy, developing call center technology that deliberately “neutralizes” the accents of Indian and Filipino staff members, as Bloomberg reported in March.
The company positions this as combating “accent discrimination” that workers face when callers struggle to understand them.
Endangered languages and dialects
The digital age poses significant threats to linguistic diversity and regional dialects.
“Among the over seven thousand languages still spoken today, nearly half face extinction according to UNESCO; approximately one-third maintain some online presence; fewer than 2 percent receive Google Translate support; and according to OpenAI’s internal testing, only fifteen languages—a mere 0.2 percent—achieve above 80 percent accuracy with GPT-4,” documents Karen Hao in her book “Empire of AI.”
“Language models are systematically homogenizing human speech,” confirms AI expert Henry Ajder, who provides advisory services to governments and technology companies, including Synthesia.
However, as these technologies become more sophisticated, they simultaneously become more potent tools for malicious actors.
Synthesia’s Express-Voice will launch as a premium service in the coming weeks, incorporating robust safeguards against hate speech and explicit content generation.
Yet numerous free, open-source voice-cloning applications remain readily available with minimal protective measures.
Early July brought reports of AI-generated messages mimicking US Secretary of State Marco Rubio’s voice being distributed to government ministers, demonstrating the technology’s potential for misuse.
“The open-source voice cloning landscape has undergone explosive development over the past nine to twelve months,” Ajder emphasizes.
“From a security standpoint, this rapid advancement presents genuine cause for alarm.”
Author: AI
Published: 15 July 2025