Nova-3 Multilingual Model Improvements

Deepgram has released an updated Nova-3 multilingual model focused on enhancing real-world multilingual speech recognition. The update delivers accuracy improvements across all supported languages, with the most significant gains in code-switching scenarios where multiple languages are mixed within a single utterance or conversation.

Key Improvements

• Lower Word Error Rate (WER): Both batch and streaming inference show improved accuracy across all languages supported by the multilingual model
• Enhanced code-switching handling: Reduced word drops when languages are mixed, addressing a common challenge in real-world multilingual applications

What Developers Need to Know

These improvements are available immediately without requiring any changes to your existing API calls or configuration. The update is backward-compatible, so you can start benefiting from the improved accuracy right away.

For additional details on supported languages and model specifications, see the Models and Language Overview documentation.