Deepgram converts speech to text (and text to speech) through an API, billed per minute or per character with no subscription, and a $200 free credit to start with no card required. Pick it when you're transcribing audio at scale and want a straightforward per-minute price; check a specific model's accuracy on your own audio before committing to volume.
Other good fits
A real-time application that needs streaming speech recognition, not just batch transcription.
+ Pros
Read pros
Deepgram is a hosted speech API: send it audio and get text back, either from a pre-recorded file/URL or as a live stream. It also offers text-to-speech (Aura and Flux TTS models). Pricing is pay-as-you-go with no subscription and no minimum, and a $200 free credit is available with no credit card required to start.
Pick it when…
You're transcribing a real volume of audio and want a predictable per-minute price rather than a subscription tier.
You need streaming (real-time) speech recognition for a live application, not just after-the-fact transcription.
You want to try a specific model's accuracy on your own audio before committing, using the free credit.
− Cons
Read cons
Accuracy varies by audio quality, accents and background noise, so a model's advertised rate says nothing about how well it'll do on your specific recordings; test on real (or realistic) audio before committing to volume. It's a hosted API, so transcription requires sending audio to Deepgram's servers; there's no offline mode.
Pick something else when…
You need transcription to run fully offline. A local Whisper runtime keeps audio on your own machine, at the cost of running it yourself.
You want speech generation, not recognition, as the primary product.Cartesia or ElevenLabs focus specifically on text-to-speech.
You want transcription as one step in a larger automated workflow. Wire the API into n8n instead of calling it directly from your own code.
Avoid if
You only need to transcribe a handful of short clips occasionally; the $200 free credit likely covers that without ever needing a paid plan.
Connection
Send audio to Deepgram's API and select a model: Nova-3 for general transcription (monolingual or multilingual), Flux for streaming, or Whisper Large as an alternative, in either pre-recorded or streaming mode. Text-to-speech works the other way: send text, choose an Aura or Flux TTS voice, and get audio back.
Pay As You Go has no minimum, no credit expiration and no card required to start with the $200 free credit. A Growth plan (pre-paid annual credits, $4K+/year) saves roughly 13–20% on these rates for higher-volume use. There's nothing to install; the ongoing effort is watching per-minute or per-character spend as volume grows.
Data handling
Audio and text sent to Deepgram's API is processed on its hosted infrastructure; check current data-retention and processing terms for anything containing sensitive recordings, since those specifics weren't independently verified beyond the pricing page for this review.
First request
Documentation-based suggestion — not run here.
Transcribe a short recorded conversation (a made-up interview or a public recording you're allowed to use) with Nova-3, then read the transcript against the audio and count the errors, paying particular attention to names and numbers, since those are where transcription tools most often slip.
Done when you've checked at least five minutes of audio against its transcript and can say whether the error rate is low enough for your use case (searchable notes) or too high (a verbatim legal record, for instance).
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Rests on Deepgram's pricing page and its recorded-audio and model-options documentation, read on 2026-09-23. No audio was transcribed and no API call was made for this review.
Deepgram pricing · Sep 23, 2026 — Pay As You Go ($200 free credit, no card), Growth plan, per-minute STT and per-character TTS rates.
Recorded audio quick start · Sep 23, 2026 — Public sample URL, Token authentication, Nova-3 and smart-format request.
Speech model options · Sep 23, 2026 — Nova-3 variants and distinction from Flux conversational recognition.
Further review notes & alternatives
Alternatives and trade-offs
Cartesia wins when you need text-to-speech as the primary product, not transcription.
ElevenLabs wins as a closer comparison for voice generation quality specifically.
n8n wins when transcription should be one step in a larger automated pipeline rather than a direct API integration.
Transcribe recordings in batches or caption live microphone streams.
Choose it when technical vocabulary and noisy audio need speech-to-text processing.
Its API supports both prerecorded and real-time application workflows.