Deepgram Nova-3

À propos de Deepgram

Fondée 2015 · San Francisco, California, USA

Deepgram was founded in 2015 by Scott Stephenson (CEO) and Noah Shutty after they finished PhDs in dark-matter physics at the University of Michigan and built a Hadoop pipeline to search audio archives. The company pivoted to commercial speech recognition in 2018 and became one of the first speech vendors to ship a fully end-to-end deep-learning ASR pipeline rather than the classical Kaldi stack. Deepgram has raised over $86M in venture funding from Tiger Global, Y Combinator, Wing, Madrona and NVIDIA's NVentures, with a 2024 Series C at a reported $1B+ valuation. The Nova model family launched in 2023 (Nova-1), followed by Nova-2 (2024) and Nova-3 (January 2025), positioning Deepgram as the fastest commercial ASR provider with sub-300 ms streaming and self-hostable deployments.

Architecture

End-to-end encoder-decoder speech-to-text with self-supervised pretraining

Deepgram Nova-3 is an end-to-end deep-learning automatic speech recognition model trained on what Deepgram describes as the largest commercial training corpus among production ASR providers, including 50,000+ hours of curated multi-domain audio plus self-supervised pretraining on hundreds of thousands of hours of unlabelled speech. Nova-3 introduced a unified multilingual model that natively handles English, Spanish, French, German, Italian, Portuguese, Hindi, Japanese, Mandarin, Korean and 10+ more, with on-the-fly code-switching inside a single audio file. New capabilities include keyword prompting (custom vocabulary at request time without retraining), Self-Hosted edition for HIPAA/SOC2 deployments, and improved noise robustness on call-centre and phone audio. The streaming variant runs at sub-300 ms latency over WebSocket, while the pre-recorded variant supports up to 4 hours per file. Outputs include diarisation, punctuation, smart formatting, language detection and topic/intent metadata.

Paramètres

Undisclosed (multi-billion)

Contexte

14 400 tokens

Capacités

• Real-time streaming ASR at sub-300 ms latency
• Multilingual single model with code-switching across 36+ languages
• Keyword prompting for runtime custom vocabulary
• Speaker diarisation, punctuation and smart formatting
• Phone-call optimised (8 kHz, low-bandwidth, noisy)
• Self-Hosted deployment for HIPAA / SOC2 / on-premises
• Up to 4 hours of audio per pre-recorded request
• Best for: call centres, real-time meeting transcription, voice agents, compliance recording

Entraînement et licence

Self-supervised pretraining on hundreds of thousands of hours of unlabelled audio plus supervised training on 50,000+ hours of curated, labelled multilingual speech. Data is sourced from licensed corpora, customer opt-in audio and publicly available datasets.

Licence: Proprietary commercial API and Self-Hosted licence. Commercial use is permitted; Self-Hosted requires a separate enterprise agreement.

Tests de sécurité: SOC 2 Type II and HIPAA compliant; no model-output watermarking required for transcripts; bias testing reported in the Nova-3 launch blog.