Deepgram: Speech-to-Text API Price and Accuracy

Deepgram is a cloud API that turns recorded or live audio into text. You send it a file or a stream, and it sends back a transcript with timestamps, punctuation and, if you ask, speaker labels. Its main model, Nova-3, lists at $0.258 per hour of recorded English audio on pay as you go ¹, and a new account starts with a $200 credit ², enough for about 775 hours of that audio ³.

The short verdict for a developer: Deepgram is a fast, well-priced streaming engine with flexible deployment, but not the most accurate API on public English tests. On Artificial Analysis's three-dataset index, Nova-3 scored a 5.2 percent word error rate ⁴, behind AssemblyAI's Universal-3.5 Pro at 3.0 percent ⁵. In the same firm's streaming benchmark, its Flux model returned a final transcript 0.020 seconds after speech ended, the fastest of the services tested ⁶. One more thing to know before you send real audio: requests are retained to improve Deepgram's models by default unless you opt out ⁷.

Every price below was read on Deepgram's pricing page on 27 October 2026. Prices change, and the streaming rates were marked as a limited-time promotion that day, so check the page before you budget. For the questions people ask most, see how much Deepgram costs, how accurate Deepgram is and Deepgram against AssemblyAI.

Deepgram sells speech recognition as an API for developers

Deepgram does not sell an app for typing by voice. It sells automatic speech recognition as a building block: a REST endpoint for recorded files, a WebSocket endpoint for live audio, plus text-to-speech and a voice agent API on the same account. The buyers are teams building call-center analytics, meeting tools, captioning and voice agents.

The company is well funded. It announced a $130 million Series C at a $1.3 billion valuation in January 2026 ⁸, after a $12 million Series A ⁹ and a $47 million round led by Madrona in 2022 ¹⁰. Developers use it widely: its JavaScript SDK was downloaded 3,682,639 times in the month to 26 October 2026 ¹¹, and its Python SDK 3,166,523 times ¹². AssemblyAI's Python SDK had 2,797,588 downloads in the same month ¹³. Download counts mostly measure automated builds, so read them as a sign of how much software depends on a package, not as a count of customers.

The models: Nova-3 for most audio, Flux for voice agents

Deepgram released Nova-3 on 12 February 2025 and calls it its most advanced speech-to-text model ¹⁴. The current line-up has five families, and the choice of model sets the price, the languages and the features you get.

Deepgram's own documentation is frank about the hosted Whisper option: it says these models are less scalable than all its other models, and that non-Whisper models return results faster and scale to higher load ¹⁶. Use it if you need Whisper's output specifically. Our guide to Whisper explains what the model is and the ways to run it.

You can reach Deepgram's models through other platforms too

Deepgram's models are resold by cloud platforms, sometimes at a different price. Cloudflare's Workers AI lists Nova-3 at $0.0052 per audio minute, or $0.312 per hour ¹⁷. Telnyx lists Deepgram Nova 3 at $0.005 per minute on its own API ¹⁸. These routes can make sense if you already bill through that platform, but check the reseller's own data terms as well as Deepgram's.

An hour of recorded audio costs $0.216 to $0.312, and streaming costs more

The price depends on four choices: recorded or live audio, one language or many, pay as you go or the prepaid Growth plan, and which add-ons you switch on. Deepgram lists its prices per minute of audio; the table converts each rate to dollars per hour (the per-minute price times 60).

Two points in that table are easy to miss. First, live audio costs more than recorded audio: Nova-3 Monolingual streams at a regular $0.0077 per minute, against a current promotional $0.0048 ³⁰. The page labels the streaming rates "limited-time promotional" and shows no end date, so budget at the regular rate and treat the promotion as a discount you may lose. Second, multilingual audio costs about 21 percent more than single-language audio for recorded files on pay as you go ³¹, and the other rows show a similar premium.

Pay as you go, Growth and Enterprise

Pay as you go has no minimum, no expiry on credit and no credit card requirement to start. The Growth plan asks for prepaid credit of $4,000 or more a year and, in return, cuts rates by up to 20 percent ³². Enterprise is priced through sales and is the tier where Deepgram signs HIPAA business associate agreements ³³.

Plans also set how much you can run at once. Pay as you go allows up to 50 concurrent REST requests and 150 streaming connections, and Growth raises streaming to 225 ³⁴. The pre-recorded documentation gives a different figure, up to 100 concurrent requests per project ³⁵, so confirm the limit for your account before a launch.

Add-ons can raise the bill, but the common ones are included for recorded audio

For pre-recorded audio, speaker diarization and smart formatting (punctuation, casing, dates and currency) are included in the model price. On streaming audio, diarization is an extra $0.0020 per minute ³⁶. Keyterm Prompting, which biases the model toward your product names and jargon, adds $0.0013 per minute, and redaction of personal data such as card and phone numbers adds $0.0020 per minute ³⁷.

So a live support line that needs speaker labels and redaction pays both add-ons on top of the model rate, while the same calls transcribed afterward as recorded files pay only for redaction.

What the $200 free credit buys

The $200 credit is a one-time starting balance, after which pay as you go billing begins ². Spent on Nova-3 Monolingual recorded audio, it lasts about 775 hours ³. Streaming at the regular rate or multilingual audio uses it faster. That is far more than a proof of concept needs, so you can run a real test (described at the end of this guide) before you pay anything.

Deepgram is accurate on public tests but not the most accurate

Accuracy is measured as word error rate (WER), the share of words a system gets wrong; lower is better. Three kinds of evidence exist for Deepgram, and they point the same way: Nova-3 is a solid general model, a step behind the leaders on English test sets.

Artificial Analysis's index puts Nova-3 behind the newest rivals

Artificial Analysis scores hosted APIs on three test datasets with one method, which makes it a like-for-like comparison of paid APIs. Its index gives Nova-3 from Deepgram 5.2 percent ⁴. The same index gives AssemblyAI's Universal-3.5 Pro 3.0 percent ⁵, OpenAI's GPT Transcribe 3.3 percent ³⁸, Google's Chirp 3 4.3 percent ³⁹ and ElevenLabs' Scribe v2 2.2 percent ⁴⁰.

One detail in that table is a useful warning. The same Nova-3 model, served through Telnyx, scored 4.8 percent ⁴¹, better than through Deepgram's own API. Settings, model versions and test dates differ between listings, so a gap of half a point should not decide your choice. A gap of two points, as between Nova-3 and the leaders, is more meaningful.

Deepgram does not appear on the Open ASR Leaderboard

Hugging Face's Open ASR Leaderboard is the other major public ranking, and it includes paid APIs as well as open models. In the English results we read on 27 October 2026, it ranked AssemblyAI's Universal-3.5 Pro at 4.87 percent ⁴² and OpenAI's GPT Transcribe at 5.49 percent ⁴³, but it listed no Deepgram model, so no direct comparison is possible there. Bear in mind that the leaderboard scores normalized text, with punctuation, casing and filler words removed ⁴⁴, so it says nothing about how well a system formats a transcript. Our question on what the Open ASR Leaderboard is explains how it works.

An academic study of lecture recordings put Deepgram behind Whisper, Microsoft and Amazon

A 2024 study measured eleven speech recognition services on university lecture recordings ⁴⁵. It predates Nova-3, so it shows Deepgram's earlier models, but it is the only academic comparison in this guide.

Each vendor's own benchmark favors itself

At launch, Deepgram reported a 5.26 percent median WER for Nova-3 on batch audio and 6.84 percent on streaming in its own tests, which it described as 47.4 and 54.3 percent lower than unnamed competitors ⁵¹. AssemblyAI's comparison page, on its own test, gives Nova-3 an average of 7.9 percent against 5.93 percent for Universal-3.5 Pro ⁵². Neither test audio is published, so neither figure can be checked. Our question on why speech-to-text accuracy claims differ explains how the choice of test set moves these numbers.

On live audio, Deepgram trades a little accuracy for speed

Streaming is where Deepgram stands out. In Artificial Analysis's 2026 streaming benchmark, Flux was the fastest service on both the first partial and the first final transcript after speech ended, at 0.020 seconds for the final, with a 7.36 percent WER ⁶. AssemblyAI's realtime model took 0.47 seconds to its first partial, at 4.46 percent ⁵³, and Flux's first partial had a 7.4 percent WER against 3.6 percent for ElevenLabs' realtime model ⁵⁴. For a voice agent that must answer quickly, the speed may matter more than the accuracy gap. For captions or records that people will read, it may not. The guide to speech recognition benchmarks covers both leaderboards in depth.

Streaming, languages and formatting are Deepgram's strengths

Beyond the raw transcript, Deepgram's case rests on its live-audio design, its language coverage and the features bundled with recorded audio.

Two kinds of streaming: Nova-3 for captions, Flux for conversations

Nova-3 streams over a WebSocket connection, and pay as you go allows up to 150 of them at once ³⁴. Deepgram says it delivers streaming latency under 300 milliseconds ⁵⁵, a vendor figure without a published method. Flux adds turn detection inside the model, so a voice agent learns when a caller has stopped talking without a separate voice activity detection step. Deepgram defines end-of-turn latency as the time from when a user stops speaking to when that end-of-turn event arrives ⁵⁶, which is the number that sets how quickly an agent can reply. Flux runs in English and, as Flux Multilingual, in 10 languages ⁵⁷.

Our guide to real-time speech recognition explains partial and final results and why latency is measured from the end of speech. For agent design more broadly, see the guide to AI voice agents.

Language counts depend on which page you read

Deepgram's pages give three different language counts for its Nova models, plus a separate count for Flux, and each is correct for what it counts.

The multilingual model handles code-switching, where a speaker changes language mid-sentence, and tags each word with its language. For a language missing from Nova-3, Deepgram points you to Nova-2 ¹⁵. Google Cloud's Speech-to-Text claims over 125 languages ⁶¹, so for long-tail languages it covers more ground.

Formatting, speaker labels and custom terms

Deepgram's diarization assigns a speaker to every word in the transcript ⁶². Smart formatting is Deepgram's name for inverse text normalization: it turns "twenty dollars" into "$20" and adds punctuation and capitals, and it is included in the price for recorded audio ³⁶. Keyterm Prompting, Deepgram's form of custom vocabulary, accepts up to 100 terms per request within a limit of 500 tokens, on Nova-3 and Flux ⁶³.

Limits for recorded files

Pre-recorded files can be up to 2 GB ³⁵. A request that takes more than 10 minutes to process with Nova returns a 504 timeout error, or after 20 minutes with Whisper ⁶⁴. For long video, Deepgram advises extracting the audio track first ³⁵, which also cuts upload time.

Deepgram keeps your audio to train its models unless you opt out

This is the point a developer most often misses. By default, requests take part in Deepgram's Model Improvement Program, Deepgram retains them, and retained data may be processed outside your region ⁷. Opting out is a per-request switch, mip_opt_out=true, and Deepgram says it then gives zero data retention: it does not store the audio, text or transcripts after the response is returned ⁶⁵. Request metadata and usage logs are kept and retrievable for 90 days, and Deepgram says they contain no audio or transcripts ⁶⁶. The glossary entry on zero data retention explains what the term does and does not cover.

Regional endpoints keep data in the EU, Australia or India

Deepgram runs three regional endpoints, in the European Union, Australia and India ⁶⁷. There is a catch: Deepgram says a full in-region guarantee needs both a regional endpoint and the opt-out flag ⁶⁸, because data retained for training may leave the region. For processing pinned to one country, it offers Deepgram Dedicated or a self-hosted deployment ⁶⁹.

Compliance: SOC 2, and HIPAA on Enterprise

Deepgram lists SOC 2 Type 1 and Type 2 certification and says it signs business associate agreements for Enterprise customers who handle protected health information ³³. A pay-as-you-go account should not assume it is covered. The guide to HIPAA-compliant dictation explains what a BAA requires, and our guide on whether dictation apps train AI on your voice compares training defaults across products.

Compared with AssemblyAI, the terms differ. AssemblyAI offers zero retention for streaming when the customer has opted out of training ⁷⁰, but says pre-recorded transcription does not offer it ⁷¹. Deepgram's statement of its opt-out names no such restriction to one mode. If you are not building a product and only want to dictate on your own computer without sending audio anywhere, a cloud API is the wrong tool: an on-device app such as Yap, which we make, runs recognition locally (see the download page), and our guide to on-device and cloud speech recognition sets out the tradeoffs.

Against AssemblyAI, OpenAI and Google, Deepgram wins on speed and deployment, not price or accuracy

The four APIs most developers compare are Deepgram, AssemblyAI, OpenAI and Google Cloud. On list price for recorded English audio, Deepgram sits in the middle. On accuracy, it trails the newest models of the other three on Artificial Analysis's index. Its real advantages are streaming speed, included diarization on recorded audio, regional endpoints and self-hosting.

Deepgram against AssemblyAI

These two are the closest rivals. On recorded English audio, AssemblyAI is cheaper and scored better on the independent index: Nova-3 costs about 1.23 times Universal-3.5 Pro per hour ⁸⁰, and AssemblyAI's lower-priced Universal-2 lists at $0.15 per hour ⁸¹. Deepgram includes diarization in that price, while AssemblyAI charges $0.02 per hour for it ⁸², so the gap narrows slightly when you need speaker labels. On live audio for voice agents, Deepgram's Flux was faster in the streaming benchmark, and Deepgram's own comparison page pitches itself on exactly that ground: real-time transcription, self-hosted deployment and one platform for voice agents ⁸³. Choose by workload: recorded English audio where accuracy matters most favors AssemblyAI, and low-latency conversation favors Deepgram. The guide to AssemblyAI covers the other side, and the Deepgram vs AssemblyAI question gives the head-to-head in brief.

Deepgram against OpenAI

OpenAI's GPT Transcribe costs almost the same for recorded audio, $0.27 per hour ⁷³ against $0.258, and scored better on the index ³⁸. OpenAI's older whisper-1 endpoint lists at $0.36 per hour ⁸⁴. Where Deepgram pulls ahead is live audio: OpenAI's realtime transcription model lists at $1.02 per hour ⁷⁶, against $0.462 for Deepgram's regular Nova-3 streaming rate ³⁰. Our question on OpenAI's speech-to-text API pricing has the full price list.

Deepgram against Google Cloud

Google's standard V2 rate is $0.96 per hour ⁷⁴, against Deepgram's $0.258, though its dynamic batch tier lists at $0.003 per minute ⁸⁵ if you can wait for results. Google's Chirp 3 scored better than Nova-3 on the index, and Google covers far more languages. Deepgram is the simpler choice for streaming and for teams that do not already run on Google Cloud.

For every provider side by side on the same measures, see the speech-to-text API comparison.

Test Deepgram on your own audio before you commit

Public benchmarks use audiobooks, meetings and earnings calls, not your users. The free credit is large enough to run a proper test, and the same steps work for any rival.

1. Collect 30 to 60 minutes of audio that matches production: the same microphones, accents, noise and vocabulary, including the hard cases.
2. Write or carefully correct a reference transcript for each file, and decide in advance whether punctuation and capitals count.
3. Run the files through Nova-3 (and Flux if you are building a voice agent) with smart formatting on, then through each rival's free tier.
4. Add your 20 to 50 most important names and terms with Keyterm Prompting and run the files again, to see how much they help.
5. Score each output with word error rate against your references, and time responses from your own servers, not the vendor's dashboard.
6. Set mip_opt_out=true on every test request that contains real customer audio, and check your logs to confirm it was applied.
7. Price a month of your real volume at the regular rates, with the add-ons you need, before you look at any promotion.

Deepgram is a strong pick when live latency, speaker labels on recorded audio, regional processing or self-hosting matter to you. If the lowest error rate on recorded English audio is the goal, test AssemblyAI and OpenAI alongside it. Our overview of speech recognition covers the wider subject, and the guide to how accurate speech recognition is puts these error rates in context.