AssemblyAI: Speech-to-Text API Price and Accuracy

AssemblyAI is a cloud API. You send it a recording or a live audio stream, and it returns a transcript, with optional speaker labels, language detection and other features. Its lowest-priced batch model, Universal-2, lists at 0.15 dollars per hour of audio ¹. Its newest batch model, Universal-3.5 Pro, lists at 0.21 dollars per hour ². The free tier includes up to 185 hours of pre-recorded transcription and up to 333 hours of streaming ³, and our question on whether AssemblyAI is free explains what the free tier includes.

On public English accuracy tests, the company's top model had the lowest error rate of the systems cited in this guide. Those scores change with each model version and test set, so test on audio like yours before you commit. For prices in detail, see how much AssemblyAI costs.

Prices here are the list prices on AssemblyAI's pricing page, read on 27 October 2026. They change, so check the page before you budget.

AssemblyAI sells speech recognition and audio intelligence as an API

AssemblyAI is a speech AI company. It sells its models through an API, with SDKs for Python and JavaScript in its documentation. Its pricing page groups the products into a few families. The pre-recorded speech-to-text API transcribes files you upload or link by URL. The realtime API transcribes live audio over a stream. The Sync API handles one short clip in a single call, with no polling or job to manage. Alongside these sit a Speech Understanding API, a Voice Agent API and a Dictation API, which the company describes as built for dictation. A self-hosted option is also listed.

The main decision is the model. For recorded audio, Universal-3.5 Pro is the model the company calls its most accurate asynchronous model, and it works across 18 languages ⁴. Universal-2 is the lower-priced batch model, and it supports 99 languages ⁵. For live audio, the company offers Universal-3.6 Pro Realtime, which the pricing page presents as its most accurate realtime model for voice agents, alongside a cheaper English-only Universal-Streaming model.

Two practical points matter for a first integration. A local file must be uploaded before it can be transcribed, which is an extra step compared with passing a public URL. And the Sync API is a different product from the batch models: it is built for short clips returned in one call, not for long files.

One hour of audio costs from 0.15 dollars on AssemblyAI, and the free tier covers 185 hours

For batch transcription, the price is a flat rate per hour of audio. The table below lists the main rates on the pricing page.

Speaker diarization is an add-on. On pre-recorded audio it costs 0.02 dollars per hour ⁷. Multichannel files are billed differently from mono files. Each channel is billed separately, so a one-hour stereo file is billed as two hours of transcription ⁸. A recording with a host and a guest on separate tracks therefore costs twice what the same hour costs as one mixed track.

The chart puts the batch rates side by side. The two AssemblyAI batch models sit at the low end, the Sync API sits above them, and the Deepgram and OpenAI rows are higher. The Deepgram and OpenAI figures are explained in the comparison below.

What the free tier covers

The free tier includes up to 185 hours of pre-recorded transcription and up to 333 hours of streaming transcription ³. At the Universal-3.5 Pro list price, the 185 pre-recorded hours would cost about 38.85 dollars ⁹. That is enough to build and test a prototype. A production workload will need a paid plan, and the free tier's limits are the vendor's to change.

Keep the free tier in mind while you test. Use it to run your own audio through each candidate model, and note the hours you use, so you know how much of the allowance a full test takes.

On public English benchmarks AssemblyAI's top model had the lowest error rate, but scores change with each version

Accuracy is measured as word error rate (WER), the share of words that are wrong, as explained in the glossary entry on word error rate. Two public leaderboards score AssemblyAI against its rivals. They use different test sets, so compare numbers within each table, not across them.

Artificial Analysis publishes an AA-WER score that averages three test datasets. It ranks Universal-3.5 Pro at 3.0 percent ¹⁰, Whisper large-v3 served through fal.ai at 4.1 percent ¹¹, and Deepgram's Nova-3 at 5.2 percent ¹².

The Open ASR Leaderboard is hosted on Hugging Face. Its English short-form results file gives Universal-3.5 Pro an average WER of 5.80 percent ¹³. The same file lists Whisper large-v3 at 7.84 percent ¹⁴. Among these rows, the AssemblyAI model has the lowest average.

These numbers move between versions. The 2025 paper that introduced the leaderboard gave Whisper large-v3 an average of 7.44 percent ¹⁵, against 7.84 percent in the 2026 file. The two figures come from different versions, so they are not directly comparable. Check the version and date before you quote a leaderboard figure.

Two cautions apply to the Open ASR scores. They are computed on normalized text, where punctuation and casing are removed, along with disfluencies such as fillers, repetitions and repairs ¹⁶. Dictation has to get punctuation and capitals right, and the leaderboard does not measure them. Its test sets are also not your audio, so run a sample of your own recordings before you choose.

For live audio, the Artificial Analysis streaming benchmark measured AssemblyAI's Universal-3 Realtime Pro at 0.47 seconds to the first partial transcript after speech ended ¹⁷. Deepgram's Flux model had the fastest first partial in the same benchmark, at 0.019 seconds ¹⁸. If latency is your main requirement, measure both on your own stream. Our guide to real-time speech recognition covers streaming and latency in general.

Diarization, streaming and languages depend on which model you choose

Speaker diarization sorts a transcript by speaker. The company's documentation says that enabling it splits the transcript by speaker, so each labeled segment has a speaker ID and its text ¹⁹. It is an add-on on the pre-recorded price list ⁷. Diarization labels who spoke, but it does not say who the speakers are. The glossary entry on speaker diarization explains the difference.

Language coverage depends on the model, so read the table closely. Universal-2 covers 99 languages ⁵. Universal-3.5 Pro covers 18 languages, and the documentation says audio in another language falls back to Universal-2 when a request asks for it ⁴. The realtime Universal-3.6 Pro model supports a longer list of languages on the pricing page, so check its count there for your language. A product that must accept many languages in one stream should check which model handles each language before it commits.

For names and terms the recognizer might miss, the company offers keyterms prompting, which is listed as an add-on on the pricing page, with a price that depends on the model. The glossary entry on custom vocabulary explains the general technique, and it is worth testing with the words your users say most often.

Against Deepgram and Whisper, the choice comes down to price, latency and data terms

The three closest rivals for a developer are Deepgram, OpenAI's hosted Whisper models, and running Whisper yourself. The table compares list prices per hour of audio. The Deepgram and AssemblyAI figures are both for pre-recorded audio.

On list price, Universal-3.5 Pro is 42 percent cheaper than the whisper-1 API ²³. Compare what each vendor will bill you over a full month of your real audio, not the first free credit or the list rate for a single model.

On accuracy, the rows cited in this guide put AssemblyAI's top model first on both leaderboards. On the Artificial Analysis index, Deepgram's Nova-3 scored 5.2 percent ¹², behind AssemblyAI at 3.0 percent ¹⁰ and Whisper large-v3 at 4.1 percent ¹¹. On first-partial latency in the streaming benchmark, Deepgram's Flux model was the faster of the two ¹⁸. Those two results pull in different directions, which is why a live product should be tested on its own audio.

Whisper is a different kind of product. You can run it on your own hardware, which means paying for the machine and the engineering time instead of a per-hour price. The guide to running Whisper locally covers that route, and our question on AssemblyAI against Whisper takes the accuracy comparison further. For the Deepgram side, see our guide to Deepgram.

What happens to your audio depends on the product, the settings and the contract

How long AssemblyAI keeps audio depends on the product and the settings, and the company's own pages do not give a single answer. The security overview says that for pre-recorded audio, uploads are deleted within 24 to 48 hours and transcripts after 30 days by default ²⁴. The data retention table gives customer-uploaded audio, for accounts without a time-to-live or BAA, a default deletion start of 72 hours, which can be set as low as one hour ²⁵. The two pages disagree on the audio default, and this guide reports both rather than choosing one.

The transcript default agrees across the pages. The FAQ says final transcripts from the pre-recorded API are stored for 30 days by default and then deleted automatically ²⁶. Customers can shorten that window with a time-to-live setting in their dashboard. For customers with a business associate agreement, all transcripts from the asynchronous endpoint are deleted after the time-to-live period, which is set to 3 days by default ²⁷.

Zero data retention is narrower. For streaming, AssemblyAI says that if you have opted out of model training, it offers zero data retention of audio and transcripts ²⁸. For pre-recorded transcription, the security overview says it does not offer zero data retention, and advises a short time-to-live or deleting transcripts through the API instead ²⁹. On the EU, the company says that data under a signed BAA, or sent to its EU servers, is never used for training ³⁰. That statement covers training only. It does not say how long EU audio is kept, so read the contract for that.

Our question on whether speech-to-text APIs store audio covers the general pattern across vendors. Before you send patient, legal or other regulated audio to any cloud API, ask the vendor in writing which retention periods apply to your account and which deletions you control. If audio must stay on your own computer, a different kind of tool is needed. An on-device app such as Yap, which we make, runs speech recognition on your computer rather than sending audio to a cloud API. Cloud APIs suit teams that accept a vendor's retention terms and need the features an API offers.

How to test AssemblyAI against your own audio before you buy

A short, fair test will tell you more than any leaderboard. Use the same files for every service, so that the results compare like with like.

1. Collect 30 to 60 minutes of audio that matches your real use: the same microphones, rooms, accents and vocabulary. Include the hard parts, such as crosstalk and product names.
2. Write a reference transcript for each file by hand, or correct a machine transcript carefully. Decide in advance whether punctuation and capitals count, because the public leaderboards do not score them.
3. Run each file through the free tier of AssemblyAI, using Universal-2 and Universal-3.5 Pro, and through the free credit of any rival you are testing.
4. Calculate word error rate for each output against the reference, as the glossary entry on word error rate describes, and compare the scores on the same files.
5. Time the responses on your network, not the vendor's dashboard, and record the total you would pay for a month of your real volume at the list prices above.
6. Before you send production audio, check the retention settings on your account and get the answer to your regulatory question in writing.

Choosing AssemblyAI comes down to the model, the price and the data terms you can accept

Choose AssemblyAI when you want a managed API with strong accuracy on recorded audio, speaker labels and a realtime option, and you can accept its retention terms. Start with the free tier and the test above. Compare the total with Deepgram and a self-hosted Whisper, using the same test files for all three.

The speech-to-text API comparison puts AssemblyAI beside other providers on the same measures. For how accuracy numbers are produced, see the guide to speech recognition benchmarks. Our overview of speech recognition covers the whole subject, and the guide to how speech recognition works explains the models behind these numbers.