Deepgram Nova-3

Nova-3 is Deepgram's most advanced speech-to-text model, released February 2025. It cut streaming word error rate by 54.3% versus competitors (6.84% median streaming WER, 5.26% batch) while keeping Nova-2-class inference speed. Its headline feature is Keyterm Prompting: developers inject up to 100 domain-specific terms at inference time to fix brand names, technical jargon, and proper nouns without retraining. It handles noisy audio, overlapping speech, numeric recognition for finance/retail, and real-time PII redaction for up to 50 entity types.

The honest picture: Nova-3 is English-first at its core — multilingual and self-hosted deployments came later and monolingual models rolled out language by language through 2025-2026. It's an API-only model with no open weights, so you're locked into Deepgram's platform and pricing. For long-form transcription accuracy, independent leaderboards have favored ElevenLabs Scribe; for the absolute lowest WER on commercial APIs, Microsoft's MAI-Transcribe-1.5 has led the Artificial Analysis leaderboard.