Getting Started with Deepgram

Speech recognition priced by the minute of audio, built for developers putting it inside something else — plus one Mac app you can simply install.

Last reviewed 2026-10-14 · for Deepgram

How it works

You open Deepgram's playground in a browser, point it at a piece of audio, and a transcript comes back within a few seconds with the words laid out on screen. Change a setting, run it again, and compare. When you want it inside your own software, that same request comes from a code file instead.

What it is

Deepgram does one narrow thing at industrial scale: it listens to audio and writes down the words, usually faster than the audio took to play. It runs the trick backwards too, reading text aloud in a synthetic voice. What it does not do is keep anything — no library of your recordings, no notes app, no Friday summary.

So the real decision is what you are shopping for. If the goal is your own meetings written down and searchable, tools built for that job — Otter and tl;dv on this site — will get you there sooner. Deepgram is sold to the people who build tools like those, which is why it is billed by the minute of audio rather than by the person using it.

Before you start

You can look before committing anything. Deepgram runs a playground page in the browser, and its own getting-started documentation sends beginners there to try the service without writing code or signing up. What that documentation does not say is whether the playground takes audio of your own or only runs Deepgram's sample — check that first, because the question that matters is whether it is accurate enough on your kind of audio.

Everything past the playground assumes code: using Deepgram properly means putting an API key (a password that identifies your app, and that spends your money) into a program. If you do not write code and have no plans to, that is not a shortcoming on your side — it is the shape of this product, and the meeting-notes tools under Where to next are the finished-article version of the same idea.

Signing up is free, Deepgram's documentation says it includes a starting balance of free credit, and the pricing page currently says no credit card is required. Transcription is billed by the minute of audio. It is a fixed pot that drains as you use it.

Security & privacy

Your audio travels to Deepgram's servers, which it says sit in the US, and comes back as text. Its compliance position is unusually strong for a tool a beginner might meet: SOC 2 Type 1 and Type 2 certification (an independent audit of how a company handles customer data), stated readiness for European privacy law, and healthcare paperwork on request.

Two things that are specific to Deepgram rather than to speech tools generally:

Deepgram's Terms of Service currently say it may use your content "to improve our Services... including training and testing our Models," under a licence it calls perpetual and royalty-free. No account-level switch turns this off. The documented way out is adding mip_opt_out=true to each request — so it lives in the code, must be there from the first request, and cannot be applied backwards to audio already sent.
Treat the API key as a card number rather than a password. It authorises spending, not merely access: whoever copies it can burn through the free credit and then whatever payment method the account carries, and a stolen key looks like ordinary use from Deepgram's side.

What it's good (and not good) at

Best at

Transcribing recordings in bulk — billed per minute of audio, currently fractions of a cent, with no per-person charge
Transcribing live, while somebody is still talking — which is why it sits inside voice assistants and phone systems
Sorting out who said what in a multi-speaker recording — speaker labelling is a documented setting on the same request, not a separate product

Okay at

Non-English audio — 50-plus languages transcribe, but Deepgram's docs reserve the full set of readability formatters for English, so a correct transcript in another language can still come back looking raw
Text-to-speech, priced per thousand characters — real, but not what Deepgram is known for

Avoid

Judging it on your very first transcript — smart_format is documented as defaulting to false, so untouched output comes back accurate but with no punctuation and no paragraphs
Budgeting from the headline streaming rates — the pricing page currently marks several of them limited-time promotional, and Flux text-to-speech is free only through 12 September 2026

First things to try

Look at what Deepgram returns before anything formats it

Runs audio through the playground — your own file if it accepts one, Deepgram's sample if not — with smart_format left at its default, and reads the raw transcript that comes back.

Why this one: smart_format is documented as defaulting to false, so Deepgram's untouched output arrives as a single unpunctuated block. Meeting that once, deliberately, is what stops you from concluding the recognition is poor when what you are looking at is a setting nobody switched on.

Tweak: If the playground lets you supply your own file, use audio containing numbers, dates or a phone number — the parts the formatter visibly rewrites, and only for English.

Do the arithmetic before you write a line of code

Estimates your monthly bill from the hours of audio you expect to process and the per-minute rate on the pricing page.

Why this one: Hours of audio per month × 60 × the per-minute rate. Billing is by the minute (current rates are on its Launchpad page), and the per-minute rate is low enough that the free credit covers hundreds of hours rather than a weekend trial. That multiplication tells you whether you are evaluating a purchase or something that stays effectively free for a year.

Tweak: Run it twice, at the pre-recorded rate and at the streaming rate. Live transcription costs more per minute, and that gap is frequently the whole design decision.

On a Mac and not writing code? Install Saga instead

Installs Deepgram's own Mac dictation app and dictates a paragraph into it, to hear how its recognition handles your voice with no code involved.

Why this one: Saga is Deepgram's own dictation app and the only route to its speech recognition with no program in between — currently free for early access, no card. It doubles as an accuracy test: if it mishears your accent, the paid service will too.

Tweak: Dictate something dense with names and technical terms. Generic prose flatters every speech tool; your own vocabulary is the real exam.

Common pitfalls

Reading the free credit as a plan rather than a balance. How fast it disappears depends entirely on how many minutes of audio you push through.
Copying model names and settings out of an old tutorial. Deepgram ships changes most weeks and has already deprecated one speaker-labelling parameter; the changelog is the source of truth.

Where to next

The rest of the developer-services category Speech tools that need no code Apps that record and write up meetings

Want the full picture first? See the Deepgram overview for pricing, the free tier and related tools.