British AI firm aims to tackle ‘accent bias’ in synthetic voices

A new voice-cloning tool from a UK firm claims it is challenging the pervasive American or “standard” southern English sound that dominates AI-generated speech.

Currently, AI voice products are trained on datasets heavily biased towards North American or Southern English speakers. This results in a monotonous, homogenized artificial voice that often sounds out of place and fails to connect with diverse audiences.

Enter Synthesia, a British company that’s spent a year building its own database of diverse UK voices. By recording individuals in studios and gathering authentic online material, Synthesia has trained its new product, Express-Voice, which can either clone a real person’s voice, preserving their unique regional accent, or generate a synthetic voice with specific local characteristics.

“If you’re the CEO of a company, or if you’re just a regular person, when you have your likeness, you want your accent to be preserved,” explained Youssef Alami Mejjati, Synthesia’s Head of Research told the BBC . The company envisions Express-Voice transforming everything from training videos and sales support to presentations, offering genuinely localized content.

While Synthesia aims for accent preservation, other companies are taking a different route. US-based Sanas, for instance, is developing tools to “neutralize” the accents of call center staff, aiming to reduce perceived “accent discrimination” by making voices sound more “standard.”

However, the rapid advancement of AI voice technology also brings new risks. While Synthesia’s product will have safeguards against misuse, the proliferation of free, open-source voice-cloning tools raises serious concerns.

The ease with which voices can now be faked for malicious purposes, as recently seen with AI-cloned voices impersonating public figures, underscores the urgent need for robust ethical guidelines and security measures in this evolving landscape.