Deepgram adds Hebrew, Persian and Urdu to its speech-to-text model
The three languages are available in production through the AI speech recognition company's Nova-3 API
Deepgram, the AI speech recognition company, has added Hebrew, Persian and Urdu to its Nova-3 speech-to-text model, with all three languages now available in production.
The release is aimed at call centres, voice agents, media transcription and analytics, and delivers what the company described as Nova-3 level accuracy and production-grade performance.
Hebrew is spoken by more than 10 million people worldwide, Persian (Farsi) by an estimated 130 million and Urdu by more than 230 million.
All three are offered as monolingual speech-to-text models designed for real-world conversational transcription in production voice environments.
Each language supports the company's Keyterm Prompting feature, which is applied dynamically at inference time to guide transcription toward domain-specific terminology, brand names and keywords without retraining models or managing custom vocabularies.
The languages are accessible through the existing Nova-3 API (application programming interface) with support for real-time streaming and batch transcription, and can be tested in the Deepgram Playground.
Related reading
- Human API reveals man-bites-dog approach to agentic AI
- Robots are leaving the lab. 2026 is when they start working for real
- End-to-end speech models cut voice latency by up to 70%, Deepgram says
Supported language codes are he, fa and ur.
Deepgram said new accounts receive $200 in credits, which the company said is enough to power more than 750 hours of transcription or 200 hours of speech-to-text across Nova-3's language suite.
The Recap
- Deepgram adds Hebrew, Persian and Urdu to Nova-3.
- Supports streaming and batch transcription via the Nova-3 API.
- Includes $200 in credits covering over 750 hours of transcription.
