Back to Gap Map Model components / Speech & audio

Deepgram Nova

closed / Overall score: n/a

Deepgram's hosted speech recognition models; Nova-3 adds real-time multilingual transcription with code-switching across 10 languages and real-time redaction. Offered as a cloud API and as a self-hosted deployment for enterprises, with no public weights.

A closed comparator. Deepgram's Flux conversational model for voice agents is a separate surface.

Openness

1 high confidence

1.0

weights: closed
data: closed
code: closed
license: Proprietary(API only)

Nova is served through Deepgram's API or a licensed self-hosted deployment; the weights are not published and no training data or code is released.

https://deepgram.com/learn/introducing-nova-3-speech-to-text-api recorded 2026-09-27
Launch post describes Nova-3 as an API model available in cloud and "Self-hosted" deployments for enterprise customers; no weights, data or code are published.

Adoption

not assessed

Nothing to count: Deepgram publishes no request, developer or customer count on its homepage or the Nova-3 launch post, and the model has no public download channel.

https://deepgram.com/ recorded 2026-09-27
Homepage offers the API to "developers and product teams" and names customers, with no usage count.
https://deepgram.com/learn/introducing-nova-3-speech-to-text-api recorded 2026-09-27
Launch post describes the model and its benchmark; no usage figure appears.

Capability

3 low confidence

3.0

Deepgram has the highest median speed factor on the board, but its word error rate trails most current commercial recognizers.

https://artificialanalysis.ai/speech-to-text recorded 2026-09-27
Table row "Nova-3" (Deepgram) with Word Error Rate "5.2%" and Median Speed Factor "564.1".

Verified 2026-09-27