Deepgram launches Arabic speech-to-text model with broad dialect coverage
Nova-3 Arabic offers transcription across 17 regional variants for production voice applications
Deepgram has released a monolingual Arabic speech-to-text model on its Nova-3 platform, designed for use in production voice applications and conversational audio, the company said.
The model supports a wide range of Arabic dialects from regions including the Middle East, the Gulf and North Africa, and achieved the lowest word error rates in benchmark tests on evaluated conversational dialects, according to Deepgram.
Nova-3 Arabic covers 17 regional variants, including Pan-Arab or Modern Standard Arabic, Gulf, Levantine, Egyptian, Maghrebi and Mesopotamian groups, as well as dialects spoken in countries such as Chad and Iran.
Related reading
- Kyle Hanagami on editing with Final Cut Pro
- Ads Decoded podcast urges marketers to refocus on strategy ahead of 2026
- Kyle Hanagami spotlights Final Cut Pro in Apple rollout of AI-powered editing tools
Deepgram said the model also supports Keyterm Prompting across dialects, enabling developers to steer transcription towards domain-specific terms at inference time without retraining the model or handling custom vocabularies.
The model is available through Deepgram’s cloud application programming interface or as a self-hosted deployment. All variants are accessible through a single API and can be trialled in the Deepgram Playground, with promotional credits available for new users.
The Recap
- Deepgram launched a monolingual Arabic Nova-3 STT model.
- Model supports 17 Arabic language variants across major regions.
- Available via cloud API or self-hosted deployment options.
