Speech-to-Text APIs Compared: Price, Accuracy, Speed

A speech-to-text API is a web service that runs automatic speech recognition on audio you send it and returns a transcript. You pay by the minute or the hour of audio. On list prices, the cheapest model we checked costs about ¹ per hour, and the most expensive model in our table costs about ² per hour. Price is only one difference. Vendor accuracy claims cannot be compared directly, independent tests show that the strongest models are close on clean English, and the rest of the choice turns on latency, speaker labels, custom vocabulary and data terms.

This guide compares the cloud platforms, the speech specialists and OpenAI's transcription models on dated list prices and independent benchmarks. For the technology behind them, read our overview of speech recognition. Short answers to the questions people ask most are on their own pages, including whether there is a free speech-to-text API, whether Google Speech-to-Text is free, how to use the Google API, what OpenAI's speech-to-text API costs and whether Gemini has a speech-to-text API.

A speech-to-text API returns a transcript for audio you upload or stream, and bills by the hour

Most providers offer two ways to send audio. In batch mode you upload a file and receive the text when it is ready. In streaming mode you send audio as it is spoken and receive partial text within a fraction of a second, which is what live captions and voice agents use. The two modes are often priced differently, and they do not produce the same quality, as the sections on speed and accuracy show.

Providers also bill on different units. Google and OpenAI quote a price per minute. AssemblyAI and ElevenLabs quote a price per hour. We converted every price to an hour so the rows can be compared. Azure bills per audio hour as well, but its public price page did not show the rates in the text we could read, so this guide does not list an Azure price.

Free tiers and trial credits are useful for a test, but they are not a production price, so the comparison below uses published list rates.

List prices for one hour of audio range from $0.15 ¹ to $1.44 ²

The table puts the list price of each model on one scale. Every row is a published rate, converted to dollars per hour of audio.

Three cautions apply to these numbers. First, they are list prices before contract discounts, and Google's pricing page also lists discounted consumption models for committed use. Second, the Google V2 figure is for the first tier of 500,000 minutes a month ¹⁰, and volumes of 2,000,000 minutes or more are listed at $0.004 per minute ¹³. Third, the Google V1 figure assumes the hour falls after the first 60 free minutes of the month ¹⁴. The rows therefore show what a small customer would pay at the published rate. They do not show what a large customer would pay.

Read the table as a range rather than a ranking. The lowest and highest rows are the same kind of product, a transcription model billed by the hour, so the gap is real. The middle rows are close enough that accuracy and features should decide between them.

Google's price depends on whether you let it log your audio

Google sells the same V1 standard recognition in two versions. With data logging the list price is $0.016 per minute after the first 60 minutes ¹⁵. Without data logging it is $0.024 per minute ¹². Opting in cuts the list price by about one third ¹⁶.

Logging is opt-in. Google's documentation states that by default Cloud Speech-to-Text does not log customer audio or transcripts ¹⁷. Under the program, Google may use logged data to refine the service, and in return you receive the lower price. The data logging documentation sets out the terms, and you should read them before you opt in, because the decision is about who may use your recordings, not just about cost. For the wider question of how voice recordings are kept and used, see voice data retention.

The V2 API is Google's newer product, and its standard rate is $0.016 per minute ¹⁰. The same price list also has a separate dynamic batch rate for non-urgent files. Check which API version your code calls, because the V1 and V2 price tables are separate.

OpenAI's transcription models are priced per minute and do not train on API data by default

OpenAI's smaller transcription model, gpt-4o-mini-transcribe, lists at $0.003 per minute ³, or about $0.18 per hour ⁴. The larger gpt-4o-transcribe lists at $0.006 per minute ⁷, or about $0.36 per hour ⁸. OpenAI's price list also shows gpt-transcribe at $0.0045 per minute ⁷, and the table above leaves that model out.

On data, OpenAI says that data sent to its API is not used to train or improve its models, unless the customer opts in ¹⁸. The same platform keeps abuse monitoring logs for all API features, and says it retains them for up to 30 days by default ¹⁹. That is a different policy from the ChatGPT apps, so check the terms for the product you use. The question on OpenAI's speech-to-text pricing goes through the billing rows in more detail.

The big cloud platforms sell breadth: languages, regions and compliance

Google, Microsoft, Amazon and IBM sell speech-to-text as one part of larger cloud accounts. Their advantages show up in language coverage, regional hosting and compliance paperwork rather than in a single benchmark.

For real-time transcription, Microsoft documents that audio is processed only in server memory and that no data is stored at rest ²⁰. Azure also lets you bias recognition toward your own words with a phrase list, and you set the list's weight within a range of 0.0 to 2.0 ²¹. That feature is the cloud equivalent of custom vocabulary in a desktop dictation app.

We did not record list rates for Amazon Transcribe or IBM Watson. Amazon's pricing page shows its rates in tiered examples, and the examples do not state a single tier that a reader can use as a list price, so we left them out instead of guessing.

Speech specialists price by the hour and publish their own accuracy claims

AssemblyAI and ElevenLabs publish their prices per hour and make accuracy claims on the same pages.

AssemblyAI lists its Universal-2 model at $0.15 per hour ¹ and describes it as supporting 99 languages ¹. Its newer Universal-3.5 Pro model lists at $0.21 per hour ⁵, and the page calls its speaker diarization "our most accurate speaker diarization yet", a claim the page does not test. ElevenLabs lists Scribe v2 at $0.22 per hour ⁶ and its real-time version at $0.39 per hour ⁹, so the real-time version lists at a higher rate than the batch version of the same model family.

For clinical use, read the HIPAA-compliant dictation guide before you rely on any vendor's medical claims.

Deepgram also publishes per-minute prices by model, but its pricing page shows promotional and regular rates side by side, so this guide does not list a single Deepgram rate.

Vendor accuracy claims are not comparable, so use independent benchmarks and your own audio

Vendors test on their own data with their own text rules, so their accuracy numbers differ in ways that are hard to see. Independent work gives a better guide, and the most useful public benchmark is the Open ASR Leaderboard, which compares open-source and proprietary systems on shared test sets. On its English short-form track, the two systems in the table below are close to each other.

The leaderboard's word error rate is measured on normalized text. Punctuation and casing are removed, and fillers, repetitions and repairs are removed too ²⁴. That is the right choice for comparing recognition, but it means the scores say nothing about punctuation, capitalization or whether a dictation tool keeps the words "um" and "you know". Those are the parts of a transcript that most people edit by hand.

Results also move with the speaker and the recording. A 2025 study of read, non-native English found that Whisper reached a mean match error rate of 0.054 and AssemblyAI 0.056 ²⁵. The two scores are very close on that one corpus. A 2024 study of eleven common services on university lecture recordings found that accuracy varied widely between vendors and between individual recordings ²⁶.

Speech that is far from typical can break every system. A 2026 study of commercial systems found that every one exceeded 51 percent word error rate on severe dysarthria ²⁷. If your users include people with atypical speech, test on their recordings, not on a benchmark. The accuracy guide explains how these numbers are built.

Open models also have a failure mode that matters for dictation. A 2024 study found that about 1 percent of Whisper transcriptions contained entire hallucinated phrases or sentences that were not in the audio ²⁸. Check for invented text in your own tests, not just for word errors. For a fuller account of how word error rate is calculated, see the word error rate guide and the glossary entry.

Speed depends on streaming or batch mode and on the decoder, not only on the model name

Speed has two parts: how quickly the system finishes a file, which the Open ASR Leaderboard reports as inverse real-time factor, and how quickly words appear while someone is still talking. The leaderboard's own analysis groups the two. Its authors report that token-and-duration transducer (TDT) decoders offer superior inverse real-time factor, which makes them better suited for long-form and batched processing ²⁹. In the table above, that split shows up as a Parakeet TDT model at 3,390 against Whisper Large v3 at 146 ^{22 23}. A higher inverse real-time factor means the system processes audio faster than real time by that multiple.

Streaming is a different test. A 2024 study found significantly lower quality for streaming speech recognition, the mode used for live events ³⁰. If your use is live captions or a voice agent, measure the streaming endpoint itself rather than the batch result. For the mechanics, see the guide to real-time speech recognition and the glossary entry for the real-time factor.

Speaker labels, punctuation and custom vocabulary often decide the choice

Headline price and accuracy are the first things to compare, but the features you turn on can change the final bill and the editing work. Three features matter most for transcripts people read.

Punctuation is often off by default. Google's documentation says that Cloud Speech-to-Text leaves punctuation marks out of recognition results unless you request automatic punctuation ³¹, so a test that uses default settings will look worse than the product you would ship. Check the formatting options in every provider before you compare text, because the normalized scores in the benchmark ignore them.

Speaker labels, which is what the industry calls speaker diarization, are often a paid add-on or a separate model. AssemblyAI's pricing page treats speaker diarization as part of its newest model and calls it "our most accurate speaker diarization yet", a vendor claim we could not test here ⁵. Ask each provider for diarization on your own two-speaker recording and count the speaker changes it gets wrong.

Custom vocabulary matters for names, product terms and medical or legal words. Azure's phrase list is the clearest documented example, with a weight range from 0.0 to 2.0 ²¹. Most providers now offer a similar keyterm or boosting feature, sometimes as a paid add-on. Test it with your own terms, because a boost that helps one word can change how others are written. For the desktop equivalent, see custom vocabulary.

Timestamps are the last feature to check. If you need word-level times for subtitles, confirm that the output includes them before you build around the API. The timestamps guide covers how they are used.

Data retention and training terms differ by provider and by plan

For many buyers, data terms decide the question before price does. The providers in this guide differ in three ways: whether they train on your audio by default, how long they keep logs, and whether they offer a zero-retention option.

The table is a summary of each provider's own statements, not legal advice. Read the data processing terms for the exact product and plan you will use, and check whether the terms change when you move from a trial to a paid account. If your recordings are sensitive, the guide on whether voice to text is safe explains the questions to ask.

The strongest privacy option is to avoid sending the audio at all. Open models can also run on your own machine, which the next section covers. Yap, which we make, is an on-device dictation app for Mac and Windows, so it is built around the same choice for people who do not want their audio sent to a cloud service.

Self-hosting an open model can be cheaper at volume, but it moves the cost to hardware and staff

An open model removes the per-minute bill, but it adds hardware and maintenance. The large Whisper model needs about 10 GB of video memory, according to OpenAI's model table ³². A machine with that much memory, the power to run it, and an engineer to keep it current is a fixed cost.

Whether self-hosting is cheaper depends on your monthly minutes, the hardware you already own and how many people maintain the service. We did not measure that break-even point, so this guide does not give one. The guide to running Whisper locally and the open-source models guide cover the setup in more detail.

Test three providers on 30 minutes of your own audio before you commit

Benchmarks rank systems on someone else's recordings. Your own recordings decide whether a system is good enough for you. A test takes about an afternoon and needs nothing beyond the providers' free trials and a spreadsheet.

1. Choose 30 minutes of real audio. Include the microphone, background noise, accents, names and the technical terms you use, because those are where systems differ.
2. Transcribe it with three providers on the same settings. Turn on punctuation and speaker labels if you will use them, and keep the custom vocabulary setting the same across providers.
3. Write a reference transcript for at least five minutes of the audio, by hand or from a careful edit of one provider's output.
4. Normalize both texts the way the Open ASR Leaderboard does, by removing punctuation and casing ²⁴, or keep punctuation if your product needs it. Compare the same rules across providers.
5. Calculate the word error rate for each provider, using the word error rate guide, and count any invented sentences separately.
6. Multiply the minutes by the list rate for the amount you expect to process each month. Use the per-hour figures in the table above, and divide the minutes by 60 first.

Keep the provider with the lowest error rate that fits your budget, your data terms and your latency needs. If two providers are close on accuracy, choose on price and terms. A cheap model that invents text or reads your names wrong costs more in editing than it saves in billing.

For a one-off file, free tools may be enough. Our browser-based file transcription is free, and online dictation works in the browser for short notes. Those are better for a single recording than for a production integration with an API key and a billing account.