What is the cheapest speech-to-text API?

If you compare only list prices, the cheapest hosted model we checked is AssemblyAI's Universal-2, and the next model, OpenAI's gpt-4o-mini-transcribe, is close behind. A low list price is not the same as a low bill, though. Free allowances, add-ons and self-hosting all change what you pay, and a model that is cheap but inaccurate for your recordings costs you in editing time. The speech-to-text API guide has the full price table, and the question on which API is best covers the accuracy side.

Universal-2 has the lowest list rate we checked

AssemblyAI lists its Universal-2 model at $0.15 per hour of audio ¹. OpenAI's gpt-4o-mini-transcribe lists at about $0.18 per hour ², a little more. Google's V1 standard recognition without data logging sits at the other end of our table, at about $1.44 per hour ³.

These are published list rates before any contract discount. Each provider bills in its own unit, so we converted every price to an hour for the comparison. Amazon Transcribe is not in the table, because its pricing page gives tiered examples rather than one rate we could use.

Free allowances beat a bill for small volumes

Free tiers change the answer if your volume is small or occasional. AssemblyAI's free tier includes up to 185 hours of pre-recorded transcription ⁴. Amazon Transcribe's free tier covers up to 60 audio minutes a month ⁵, for the first 12 months after signup. Free tiers are for a trial, not a production price, so check the terms before you rely on one for a service that runs every day.

Self-hosting has no per-minute fee but a hardware cost

Open models carry no per-minute fee. The Whisper code and model weights are published under the MIT License ⁶, so you can run them yourself. The cost moves to hardware, power and the time of whoever maintains the setup. As a rough break-even against AssemblyAI's rate, a rented GPU server costing about $383.98 a month ⁷ would need to process about 2,560 hours of audio a month ⁸. Below that volume the hosted rate is cheaper on paper, before you count your time.

Self-hosting makes most sense when privacy or very high volume is the reason for it, rather than the lowest bill at small scale. The guide to running Whisper locally covers the setup.

Check the add-ons before you compare

The headline rate is not the whole price. AssemblyAI charges extra for speaker labels, at $0.02 per hour of audio ⁹, so a transcript that needs speakers costs more than the base rate. Google's price also changes with data logging, which is covered in whether Google Speech-to-Text is free. For speaker labels across providers, see which speech-to-text APIs can identify speakers.

What to do: multiply your hours by the rate, then test the cheapest two

1. Total your audio for a typical month, in hours. Convert minutes by dividing by 60.
2. Multiply the hours by the list rate for each model, then subtract any free allowance that applies.
3. Add the add-ons you need, such as speaker labels or real-time streaming, because they often change the order.
4. Test the two cheapest models on 30 minutes of your own audio, using the word error rate guide. A low error rate and a low price are both needed.
5. Choose the cheapest option that meets your accuracy and data terms. If the cheapest one invents text or misreads your names, the editing cost will outweigh the saving.