What it is
Nova-3 was released by Deepgram on 12 February 2025 and remains its general speech-to-text model for recorded and streaming audio. It transcribes 63 language entries, switches between 10 of them mid-sentence in real time, and supports keyterm prompting for up to 100 domain words, speaker diarization, redaction and smart formatting. Medical and pharma variants cover English.
Deepgram has not replaced it. Its newer model, Flux, is a separate streaming model for voice agents that adds turn detection; the docs still describe Nova-3 as the top general-purpose model without turn detection, and keep Nova-2 for languages Nova-3 lacks and for filler-word detection. Enterprise customers can run Nova-3 in self-hosted containers on NVIDIA GPUs, in their own VPC or on premises.
Our review
Nova-3 is a model for throughput, not accuracy. Artificial Analysis measures it at 546 times real-time on 10-minute clips, second only to NVIDIA's Parakeet TDT 0.6B V3 on Together AI at 618, for $4.30 per 1,000 minutes. Its word error rate on the same test, 5.2%, sits near the bottom of 62 entries. ElevenLabs Scribe v2 scores 2.2% and Microsoft's MAI-Transcribe-2 2.0%.
Deepgram's own figures look far better: a median error rate of 5.26% on recorded audio against 10% for the next-best rival, and 6.84% against 14.92% when streaming. Those come from its own 2,703 files, about 82 hours across domains such as air-traffic control, drive-thru and voicemail, with the competitors unnamed. The gap is a reminder that Nova-3 was tuned for noisy phone audio, while neutral boards reward clean speech.
The case for it is the work around live audio. It switches between 10 languages mid-call, redacts up to 50 entity types in real time, and takes up to 100 keyterms for product names. Each of those is billed separately when streaming. On pay-as-you-go, diarization and redaction cost $0.0020 a minute each and keyterms $0.0013, so a fully featured stream at the regular rate runs to $0.0130 a minute, about 70% above the headline price.
The data default deserves attention before a pilot. Requests join Deepgram's Model Improvement Program unless each one carries mip_opt_out=true, and Deepgram keeps "fractional increments" of that data for training. It cites SOC 2 Type 2, HIPAA, PCI DSS and GDPR, but signs HIPAA business associate agreements only for Enterprise customers, so a clinic trialling nova-3-medical on a credit card has neither by default.
Nova-3 now sits awkwardly in Deepgram's own range. Voice-agent builders get turn detection from Flux at $0.0065 a minute on promotion, and teams transcribing archives can buy Scribe v2 for $3.67 per 1,000 minutes on Artificial Analysis's figures, less than Nova-3's $4.30 with less than half the errors. What remains is the live, high-concurrency, sometimes on-premises transcription that Deepgram built its business on.
Where it earns its place
A platform team at a contact-centre software vendor
Transcribe 1M minutes of live calls a month for agent assist and QA
Streaming costs $7,700 a month at the regular pay-as-you-go rate or $4,800 on the current promotion; adding diarization and redaction adds $4,000. Pay-as-you-go allows 150 concurrent streams and Growth 225, so a busy floor needs an Enterprise contract.
A developer building a phone-ordering bot for a 40-site restaurant chain
Catch menu items and promo codes reliably in noisy drive-thru audio
Keyterm prompting takes up to 100 terms for $0.0013 a minute extra. Nova-3 has no turn detection, so the bot needs its own end-of-speech logic, or the team moves to Flux, which builds it in.
An IT lead at a hospital group with data that cannot leave its network
Transcribe clinician dictation on premises
nova-3-medical runs in Deepgram's self-hosted containers on NVIDIA GPUs inside the hospital's VPC. HIPAA agreements, self-hosting and custom pricing all come through an Enterprise contract, not the self-serve plans.
A media company's archive manager
Batch-transcribe 100,000 hours of back-catalogue video for search
Six million minutes of recorded audio cost $25,800 on pay-as-you-go or $21,600 at the Growth rate, with diarization included. If the transcripts will be published, Scribe v2's lower error rate saves editing time for a similar bill.
Before you commit
- Setup
- Sign up for $200 of free credit without a card, which Deepgram says covers about 43,000 minutes; one REST call transcribes a file, a WebSocket handles live audio.
- Integrations
- REST and WebSocket APIs with SDKs for Java, JavaScript, Python, Go and .NET; nova-3-medical and nova-3-pharma variants; Flux and the Voice Agent API alongside.
- Data and privacy
- Requests join the Model Improvement Program by default; mip_opt_out=true on each request limits retention to processing. Deepgram cites SOC 2 Type 2, HIPAA, PCI DSS and GDPR, offers an EU endpoint, and signs BAAs only for Enterprise.
- Cost at scale
- 1M streaming minutes a month is $7,700 at regular pay-as-you-go rates, $6,500 at Growth rates, plus $2,000 each for streaming diarization and redaction. The same volume of recorded audio is $4,300, or $3,600 on Growth, which starts at $4,000 a year.
- Limits
- 50 concurrent REST requests and 150 streams on pay-as-you-go, 225 streams on Growth; up to 100 keyterms; no turn detection; some languages and filler-word detection still need Nova-2.
What's good
- 546x real-time on Artificial Analysis, second fastest on the board
- Real-time code-switching across 10 languages
- Self-hosted containers and an EU endpoint
- Recorded audio from $0.0043 a minute with diarization included
What's not
- 5.2% AA-WER, more than double Scribe v2's 2.2%
- Audio can be kept for training unless each request opts out
- Streaming add-ons lift the per-minute price by about 70%
- HIPAA agreements on Enterprise only
Best for
Contact-centre and live-captioning teams that need fast, high-concurrency transcription of phone audio, including on their own servers.
Alternatives
Deepgram Nova-3: common questions
- Is Deepgram Nova-3 worth it?
- Nova-3 is quick and cheap enough for live call audio, but on a neutral test it makes more than twice the errors of the accuracy leaders. We rate it 4.0 out of 5.
- How much does Deepgram Nova-3 cost?
- Pre-recorded $0.0043/min (monolingual) or $0.0052/min (multilingual) pay-as-you-go; streaming $0.0048/min on promotion, $0.0077 regular. Growth plan from $4,000/year is cheaper. $200 free credit.
- What are the downsides of Deepgram Nova-3?
- 5.2% AA-WER, more than double Scribe v2's 2.2%; audio can be kept for training unless each request opts out; streaming add-ons lift the per-minute price by about 70%; hIPAA agreements on Enterprise only.
- Who is Deepgram Nova-3 best for?
- Contact-centre and live-captioning teams that need fast, high-concurrency transcription of phone audio, including on their own servers.
- How much does Deepgram Nova-3 cost for a team or in production?
- 1M streaming minutes a month is $7,700 at regular pay-as-you-go rates, $6,500 at Growth rates, plus $2,000 each for streaming diarization and redaction. The same volume of recorded audio is $4,300, or $3,600 on Growth, which starts at $4,000 a year.
- Is Deepgram Nova-3 safe for company data?
- Requests join the Model Improvement Program by default; mip_opt_out=true on each request limits retention to processing. Deepgram cites SOC 2 Type 2, HIPAA, PCI DSS and GDPR, offers an EU endpoint, and signs BAAs only for Enterprise.
- What are the best Deepgram Nova-3 alternatives?
- The closest alternatives we have reviewed are ElevenLabs Scribe v2, Gemini 3.5 Transcribe, GPT Transcribe and Whisper large-v3-turbo.
Answers checked Oct 22, 2026.
Make Deepgram Nova-3? Show this rating on your site
