#Egypt #AIvoice - Alexandria-based AI lab TokenAI has released Replica Preview, a family of three open-weights text-to-speech models for voice generation and voice cloning. The models have 0.5B, 0.6B and 0.7B parameters, and the company says they generate audio quickly and cost little to run. TokenAI is working to improve Arabic speech across regional dialects. TokenAI shared the models this week, about one year after announcing the project.
SO WHAT? - Most leading speech models are paid with closed APIs, and Arabic dialects are poorly served. Over the past two years a growing variety of developers have developed Arabic models that have helped improved standards, including TTS (text-to-speech) and STT (speech-to-text) models. However, the release of three open weight Arabic speech models is significant. The open releases will allow AI teams to run the models on their own infrastructure, rather than rely on closed models in the cloud. No benchmarks or audio samples have been published yet, but TTS and STT users will be paying close attention to the development of the Replica family.
KEY POINTS:
TokenAI released the Replica Preview family this week, a year after announcing it had begun building the Replica voice models. The company calls the releases prototype models, which are its first voice releases.
Replica Preview 0.5B is a compact multilingual model that supports zero-shot voice cloning from a short reference recording. The model pairs an autoregressive speech-token model with a flow-matching vocoder, and suits lightweight and local use.
Replica Preview 0.6B is a diffusion-based, non-autoregressive multilingual model, built on OmniVoice. It handles expressive speech, zero-shot cloning, controllable voice design, sentence-chunked streaming and non-verbal sounds such as breaths, sighs and laughter.
Replica Preview 0.7B EN is an English-only model for narration, voice assistants and content production. It supports voice cloning from an authorised reference recording and is aimed at fast, real-time delivery.
The three models have all been released as open weights, so developers can run them on their own infrastructure. The models could appeal to teams that can’t justify the recurring cost, usage limits or vendor dependence of paid closed-source speech APIs.
TokenAI is working on Arabic speech generation, paying attention to dialect, pronunciation, rhythm and conversational delivery. It plans to improve dialect coverage and voice control in future versions.
The company says it believes the models can compete with, and in practical cases outperform, paid closed-source rivals on speed, deployment control and operating cost. However, no benchmark results, research data or peer reviews have yet been published to support this.
Target uses for the Replica family include voice interfaces, content production and accessibility features.
[Written and edited with the assistance of AI]
Source: TokenAI
LINKS
Replica-Preview-0.6B code (Hugging Face)
Replica Preview 0.5B (Hugging Face)
Replica Preview 0.7B EN (Hugging Face)
Read more research about Token AI:
TokenAI releases 50k reasoning records for AI builders (Middle East AI News)
TokenAI launches compact AI decision model (Middle East AI News)
TokenAI open-sources Horus Taleeq Arabic language model (Middle East AI News)
TokenAI releases Horus Cyber Nano 1.0 (Middle East AI News)
TokenAI to release Horus cybersecurity model (Middle East AI News)


