Blog / speech-to-text-api-comparison

Speech-to-Text API Comparison: Pricing and Fit for Video Teams

Quick answer

List prices for OpenAI, Deepgram, and AssemblyAI checked on 5 October 2026, plus the input, output, and pipeline differences that decide which API fits video work.

2026-10-05 | 8 min read | ReelWords Team

ReelWords illustration comparing general speech-to-text APIs with a video URL transcript API.

Most speech-to-text API comparison articles rank providers by a single accuracy number. That number rarely survives contact with your own audio, and it hides the costs that decide a video project: getting the audio out of the video, fitting under upload limits, and turning the response into subtitle files. This comparison starts with published prices, then looks at those pipeline differences.

All prices below are list prices taken from each provider's public pricing page on 5 October 2026. They change, so follow the links before you budget.

Speech-to-text API comparison: list prices

ProviderModelList pricePer hour of audio
OpenAIgpt-4o-mini-transcribe$0.003 / min$0.18
OpenAIgpt-4o-transcribe$0.006 / min$0.36
OpenAIWhisper$0.006 / min$0.36
DeepgramNova-3 monolingual (pre-recorded)$0.0043 / min$0.258
DeepgramNova-3 multilingual (pre-recorded)$0.0052 / min$0.312
AssemblyAIUniversal-2$0.15 / hr$0.15
AssemblyAIUniversal-3.5 Pro$0.21 / hr$0.21

The per-hour column is our arithmetic (per-minute price × 60) so the rows can be compared directly.

Notes on reading the table:

  • Free allowances differ. Deepgram lists a $200 free credit and AssemblyAI lists $50 in free credits on signup. A credit balance and a recurring monthly allowance are different things, so read the terms.
  • Add-ons are extra. AssemblyAI prices features such as speaker diarization and entity detection as add-ons on top of the base rate. Check which features you need before comparing base prices.
  • Self-hosting is not free. Open-source Whisper through whisper.cpp has no per-minute fee, but you pay for compute and the engineering time to run it.

We have not included Google Cloud, AWS, or Azure speech services because we did not verify their current prices for this article. They are established options and worth pricing if you already run on one of those clouds.

Why accuracy numbers do not transfer

Word error rate depends on the test audio. A model that leads on clean read speech can fall behind on phone audio, music beds, overlapping speakers, or domain vocabulary. Vendor-published benchmarks use each vendor's chosen test set, so two providers' numbers are not measuring the same thing.

The reliable method is small and boring:

  1. Pick 20 to 30 minutes of audio that represents your real input, including the hard parts.
  2. Write a correct reference transcript by hand.
  3. Run each candidate API on the same files.
  4. Count the errors that matter to you: names, numbers, and terms of art, as well as overall word errors.

An afternoon of this tells you more than any leaderboard. For the kinds of errors to count, see our guide to fixing inaccurate auto captions.

What a general speech API expects from you

Speech-to-text APIs transcribe audio. Getting from "a video somewhere" to "audio the API accepts" is your job.

  • Input. OpenAI's transcription endpoint takes a file upload, and its speech-to-text guide documents a 25 MB limit per file, so long recordings must be compressed or split. Deepgram and AssemblyAI accept a URL to a hosted media file as well as uploads. None of them resolves a social platform page: a TikTok or YouTube page URL is HTML, not media.
  • Extraction. For platform videos you need to fetch the media and pull out the audio yourself. Our guide to extracting audio from a video URL shows the FFmpeg and yt-dlp commands, along with the maintenance that comes with them.
  • Output. Responses are JSON with timestamps. Some providers can also return SRT or VTT directly; where they cannot, you write the formatter and decide line lengths and cue durations.
  • Orchestration. Long files are processed asynchronously. You need a queue, polling or webhook handling, retries, and a place to store results. The Node.js pipeline guide outlines that architecture.

None of this is hard, but it is the bulk of the work. The per-minute price is a small part of the total cost of a pipeline you have to build and keep running.

Where ReelWords fits, and where it does not

ReelWords is not a general speech-to-text API, and comparing it on per-minute price would be misleading. It is a transcript and caption API for video URLs.

What it does:

  • Accepts a public video URL from YouTube, TikTok, Instagram, X, Facebook, or a direct file link, with no audio extraction on your side.
  • Uses the platform's native captions when they exist, and speech recognition when they do not. The job reports which via transcriptSource.
  • Returns timestamped segments as JSON, and downloads as srt, vtt, txt, or txt_timestamps.
  • Renders captioned MP4s through a separate caption-jobs endpoint, using a chosen caption style.
  • Exposes the same operations to AI agents through an MCP server.

What it does not do:

  • Real-time or streaming transcription.
  • Accept raw audio uploads through the public API. The transcript endpoint takes a URL.
  • Notify you by webhook. Jobs are polled.
  • Charge per minute. Usage is plan-based, with limits and credit top-ups listed on the pricing page.

Which one should you choose?

If your situation isConsider
Live captions, voice agents, call audioA streaming speech API such as Deepgram or AssemblyAI
Large volumes of your own audio files, lowest unit costA per-minute API, benchmarked on your audio
Strict data residency or offline processingSelf-hosted whisper.cpp
Transcripts from social or hosted video URLsA URL-based transcript API such as ReelWords
Captioned video output, not only textA caption rendering API such as ReelWords

Many teams end up with two: a general speech API for their own recordings and a URL-based service for platform content. If the second case is yours, the social platform transcription API guide covers the request flow, and the Python tutorial has working code.

FAQ

What is the cheapest speech-to-text API?

On list price per hour as of 5 October 2026, AssemblyAI Universal-2 ($0.15) and OpenAI gpt-4o-mini-transcribe ($0.18) are the lowest in this comparison. Add-on features, minimums, and your own engineering time change the real cost.

Which speech-to-text API is the most accurate?

There is no general answer. Accuracy depends on your audio. Test the candidates on a representative sample with a hand-checked reference transcript.

Can a speech-to-text API transcribe a YouTube or TikTok link?

General speech APIs need audio or a direct media file URL, so you would have to extract the audio first. URL-based transcript APIs such as ReelWords accept the page link itself.

Do I need to extract audio before transcribing a video?

For APIs with upload size limits, usually yes, since audio is much smaller than video. If you use a service that accepts the video URL, no.

Is self-hosting Whisper cheaper than an API?

At low or irregular volume, usually not, once you count compute and maintenance. At high steady volume with spare GPU capacity it can be. Estimate with your own hardware costs.

Do these APIs return SRT or VTT files?

It varies. Check each provider's documentation for subtitle output. ReelWords returns SRT, VTT, plain text, and timestamped text from the download endpoint. The format differences are explained in SRT vs VTT.

How were these prices checked?

We read each provider's public pricing page on 5 October 2026 and copied the listed pay-as-you-go rate for pre-recorded audio. Volume discounts and enterprise contracts are not reflected.

Working with video URLs? See ReelWords features, compare plans on pricing, or check the FAQ.