Nova-3

Nova-3 launched on February 12, 2025 as Deepgram's most advanced speech-to-text model, going GA in 2025-Q1. It uses a latent-space architecture to encode speech patterns efficiently; Deepgram claims a 54% WER reduction versus the field, with real-time multilingual transcription, keyterm prompting for domain vocabulary, and a Nova-3-Medical variant (March 2025) tuned for clinical terminology. It became the default STT in Twilio's ConversationRelay in September 2025 and is positioned as the low-latency hot path for voice agents, often paired with Whisper as a batch fallback. Pricing is usage-based per minute via the Deepgram API. As with all vendor benchmarks, the -54% figure is team-reported; independent 2026 streaming benchmarks place Deepgram's models solidly but not at the very top on accuracy, with latency as the real strength.

Team-reported: -54% WER vs field