Alexandria-based AI lab TokenAI has released Replica Preview, a family of three open-weight text-to-speech models for voice generation and cloning.
The models range from 0.5B to 0.7B parameters. According to the company, they are designed to generate audio quickly and run at relatively low cost. Developers can deploy them on their own infrastructure instead of relying on paid, closed speech APIs.
Replica Preview 0.5B is a compact multilingual model supporting zero-shot voice cloning from a short recording. The 0.6B model is built for expressive speech, voice design, streaming and non-verbal sounds such as breaths, sighs and laughter. The English-only 0.7B model targets narration, voice assistants and content production, with support for cloning from an authorised reference recording.
The release comes about a year after TokenAI announced the Replica project and marks the company’s first voice models. TokenAI said it is also working to improve Arabic speech generation across regional dialects, including pronunciation, rhythm and conversational delivery.
No benchmark results, audio samples or peer-reviewed research have been published yet. Potential applications include voice interfaces, content production and accessibility tools.
Source: Middle East AI News


