Deepgram, the AI speech recognition company, has updated its Nova-3 Multilingual speech-to-text model to improve accuracy across supported languages, with the largest gains in code-switching scenarios where speakers mix languages within single utterances or conversations.
The company said the updated model lowers word error rate across both batch and streaming inference.
Deepgram said Nova-3 Multilingual was retrained and evaluated on diverse multilingual benchmarks, highlighting advances in curriculum design and data curation to improve the model's exposure to code-switching examples and reduce word drops when languages are mixed.
The model supports dynamic Keyterm Prompting at inference time to bias transcriptions toward domain-specific terms and names, which the company said is particularly useful for call centres, voice agents and industry analytics.
Related reading
- DeepClaw lets users create a personal AI phone agent with a single call
- 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
The updated model is live now as the default production Nova-3 Multilingual model and requires no API (application programming interface) or configuration changes, the company added.
Deepgram said new accounts receive $200 in credits, equivalent to more than 750 hours of transcription or 200 hours of speech-to-text on Nova-3.
The Recap
- Deepgram updated Nova-3 Multilingual to reduce transcription errors.
- Credits equal over 750 hours or 200 hours across Nova-3.
- The updated model is live and requires no configuration changes.
