Speech-to-text API pricing guide (2026): per-hour rates, add-ons and hidden costs

An 11-second voice command costs you 11 seconds of audio with one provider and a full minute with another. Multiply that by a million calls a month and the "cheapest" API on paper can turn into the most expensive line on your bill.

That's why headline rates don't tell you much on their own. Speech-to-text APIs differ on billing units, what's bundled, how streaming sessions are metered, and which add-ons you'll actually need. Most pricing roundups skip those details, and a lot of them still quote last year's numbers.

This guide does three things. It explains how speech-to-text APIs charge, lists current per-hour rates for the major providers in one sourced table (prices as of November 3, 2026, each linked to the vendor's own pricing page), and shows how to estimate your real cost for common workloads.

How speech-to-text APIs charge for usage

Every speech-to-text bill comes down to four variables: the billing unit, the processing mode, the model tier, and the add-ons you turn on. Get any one of them wrong and your forecast drifts.

Billing units: per second, per minute, or per block

The billing unit decides what you pay for short clips. Some providers round each request up to the nearest second, some to the nearest minute, and some to fixed blocks such as 15 seconds.

Here's what one 11-second clip costs under each method:

• Rounded to the second: you pay for 11 seconds.
• Rounded to 15-second blocks: you pay for 15 seconds, about 36% more audio than you sent.
• Rounded to the minute: you pay for 60 seconds, more than five times the audio you sent.

If your workload is long files like podcasts or meetings, rounding barely matters. If it's voice commands, IVR turns or short dictation clips, it can matter more than the per-hour rate. Check each provider's rounding rule before you compare prices. Google Cloud, for example, states that each request is rounded up to the nearest second.

AssemblyAI bills pre-recorded audio per audio hour submitted, with no minimum commitments, upfront fees or contracts on pay-as-you-go. Two rules are worth knowing before you forecast:

• Streaming is billed on session duration. You pay for the time the WebSocket is open, not the amount of audio you send. Idle time counts, so close the connection as soon as a call ends.
• Multichannel audio is billed per channel. A one-hour stereo file split into two channels is billed as two hours.

Batch, sync and streaming

Processing mode is the next biggest lever. AssemblyAI offers three, and they're priced differently because they do different jobs:

• Pre-recorded (async): you upload a file and get a transcript back when it's done. Universal-3.5 Pro is $0.21/hr, and Universal-2 is $0.15/hr.
• Sync: one HTTP POST returns a finished Universal-3.5 Pro transcript, with no polling and no WebSocket. The Sync API handles clips up to two minutes, returns results in about 134 ms at the median, and costs $0.30/hr.
• Streaming (Realtime): audio goes in over a WebSocket and text comes back while the person is still talking. Universal-3.6 Pro Realtime is $0.45/hr base, and Universal-Streaming (English or Multilingual) is $0.15/hr.

Voice agents and live captions need streaming. Meeting notes and podcast search don't, so batch is the cheaper choice there. Short clips that need a finished transcript fast, like voice commands, fit Sync. If you're still weighing the trade-offs, our guide to batch vs real-time transcription breaks it down.

Standard vs premium models

Every provider sells more than one model tier. Lower tiers handle general transcription. Premium tiers do better on proper nouns, alphanumerics like order numbers and emails, and domain terms.

The premium usually pays for itself when transcripts are shown to users, stored for compliance, or used to trigger actions like refunds or bookings. If your transcripts only feed a search index that tolerates errors, a lower tier may be enough. Our public benchmarks show where each model ranks, including where we trail, so you can judge the accuracy side of that trade-off yourself.

Bundled vs unbundled features

Providers split into two camps:

• Unbundled: base transcription plus a per-hour charge for each add-on. AssemblyAI and Deepgram work this way.
• Bundled: some features are included in the base rate. Gladia, for example, includes speaker diarization on its async plans.

Unbundled pricing is cheaper if you need only transcription and one or two extras. Here's an AssemblyAI pre-recorded example:

• Universal-3.5 Pro: $0.21/hr
• Speaker diarization: +$0.02/hr
• Keyterms prompting: +$0.05/hr
• Total: $0.28/hr

If you'd turn on five or six add-ons for every file, total them up and compare against a bundled rate before you decide.

Speech-to-text API pricing comparison (prices as of November 3, 2026)

The table below lists each provider's published pay-as-you-go list rates, taken from the vendor's own pricing page and converted to per hour (per-minute rate × 60). Volume discounts, committed-use plans and promotions aren't included.

Prices as of November 3, 2026, from each vendor's own pricing page. OpenAI lists its per-minute figures as estimated costs. Check each page before you commit, since list prices change often.

A few patterns stand out:

• Streaming carries a premium almost everywhere. It costs more to hold a live connection and return text in real time.
• Batch prices cluster tightly, streaming prices don't. Most pre-recorded rates in the table sit between about $0.12 and $0.40 per hour. Streaming rates run from $0.15 to just over $1 per hour.
• Promotions and tiers make comparisons slippery. Google's volume tiers, Deepgram's Growth-plan discounts and Gladia's EUR pricing all shift the real number. Compare on the rate you'll pay at your volume, not the lowest number on the page.
• Rate limits are a cost too. Some providers cap concurrent streams by plan. AssemblyAI has no spend commitment to unlock concurrency: pay-as-you-go accounts start at 100 new streaming sessions per minute, and that limit rises automatically as you use it.

AssemblyAI pricing at a glance (as of November 3, 2026)

Here's the full AssemblyAI lineup in one place. Everything is pay-as-you-go with no minimum commitments. For the current figures, check the pricing page.

Speech-to-text models

Add-ons for pre-recorded audio

Add-ons for streaming

Speech Understanding

Text-to-speech and LLM Gateway

AssemblyAI-TTS-1 costs $15 per 1M characters, billed on the input characters you send. Cloned voices and EU processing cost the same.

LLM Gateway is billed per million input and output tokens, with separate rates for each model. The rates on the pricing page are for global routing; in-region (US or EU) routing is 10% higher.

Processing in the US or the EU costs the same for speech-to-text. A Business Associate Addendum (BAA) is available for customers processing protected health information, with no premium pricing. See can you sign a BAA for details.

Calculating total cost of ownership

The cheapest per-hour rate isn't always the cheapest system. Your real cost depends on the features you need, how accurate the transcripts have to be, and how much engineering it takes to run the pipeline.

Cost scenarios by use case

These examples use AssemblyAI list prices so you can check the math.

Startup building podcast search, 1,000 hours a month. Search tolerates some errors, and you don't need speakers labeled.

• Pick: Universal-2 at $0.15/hr. Estimated cost: $150 a month.
• Skip: premium add-ons you won't use.

Contact center with compliance review, 10,000 hours a month. Transcripts are reviewed, stored and searched by agent and customer.

• Pick: Universal-3.5 Pro with speaker diarization ($0.21 + $0.02 = $0.23/hr). Add PII Text Redaction (+$0.12/hr) if transcripts leave the compliance team.
• Estimated cost: $2,300 a month before redaction, $3,500 with it.

Healthcare practice transcribing patient visits, 5,000 hours a month. Drug names and dosages have to be right, and you need a BAA.

• Pick: Universal-3.5 Pro with Medical Mode, $0.36/hr combined, with a Business Associate Addendum (BAA) in place.
• Estimated cost: $1,800 a month.

Voice agent handling 20,000 call-hours a month. The agent has to catch names, account numbers and turn ends in real time.

• Pick: Universal-3.6 Pro Realtime at $0.45/hr, with keyterms included. Add streaming diarization (+$0.12/hr) only if the agent needs to tell speakers apart.
• Estimated cost: $9,000 a month for streaming STT, billed on session time. For the full stack cost with an LLM and text-to-speech, see our breakdown of voice agent pricing.

Hidden costs beyond the per-hour rate

Headline rates leave out several costs that show up later:

• Idle streaming time. If you bill on session duration and leave sockets open between calls, you pay for silence. Close connections when a call ends.
• Channel count. Stereo call recordings billed per channel cost twice what a mono forecast predicts. Downmix if you don't need a transcript per channel.
• Cloud plumbing. Cloud-platform speech APIs usually sit inside a pipeline of storage buckets, functions, queues and egress, each billed separately. API-first providers take audio in and return text without that layer.
• Engineering time. Every extra service is something to configure, monitor and keep within data residency rules.
• Switching later. Migrating providers after launch costs engineering time and re-testing. Our guide on when to switch STT providers covers how to work out whether it's worth it.

How accuracy shows up on your bill

Accuracy differences become labor costs as soon as people review transcripts. If one model misses more names and numbers, your team spends more time fixing records, agents ask callers to repeat themselves, and automated actions fail. A slightly higher per-hour rate is often cheaper than the cost of corrections. Check accuracy on your own audio before you treat two providers as interchangeable.

Choosing the right pricing model for your needs

Start with your constraints, then compare prices among the providers that meet them. Data residency or a BAA requirement can rule out options before price even comes up.

Questions to answer first:

• Do you need results in real time, within a second or two, or whenever the file is done?
• Will people read the transcripts, or only machines?
• Which add-ons will you turn on for every file, and which only sometimes?
• Do you need EU processing, PHI handling under a BAA, or both?
• How many concurrent streams will you need at peak, and does the provider charge or require a commitment to get them?

Common evaluation mistakes:

• Rate shopping: picking the lowest per-minute number without adding the features you need.
• Ignoring rounding: forecasting short clips at per-second cost on a per-minute plan.
• Forgetting idle time: forecasting streaming on audio minutes when the bill counts session minutes.
• Testing accuracy late: finding out in production that the cheap model misses entities.
• Treating compliance as an afterthought: discovering a residency rule that removes your chosen provider.

The per-hour rate is only one layer. AssemblyAI runs speech-to-text, Speech Understanding, LLM Gateway, Medical Mode and text-to-speech on one platform, with one bill and no commitment to scale. Companies like Uber and ClickUp build on it.

Final words

Speech-to-text pricing now goes well past a per-minute rate. Billing units, session metering, add-ons and accuracy all move the number you'll actually pay. The provider with the lowest headline rate can become the most expensive one once you add corrections, features and engineering time.

So treat published rates as a starting point. Pick the processing mode your product needs, total the add-ons you'll actually use, test accuracy on your own audio, and then compare. And check the date on any price table you rely on, including this one.

Frequently Asked Questions (FAQs)

How much does AssemblyAI speech-to-text cost?

As of November 3, 2026, AssemblyAI's pre-recorded Universal-3.5 Pro costs $0.21/hr and Universal-2 costs $0.15/hr. Universal-3.6 Pro Realtime streaming costs $0.45/hr base with keyterms included, the Sync API is $0.30/hr, and the Dictation API is $0.62/hr. Add-ons such as speaker diarization (+$0.02/hr on pre-recorded) and Medical Mode (+$0.15/hr) are priced separately on the pricing page.

How much does AssemblyAI's streaming speech-to-text cost?

Universal-3.6 Pro Realtime costs $0.45 per hour of session time, and keyterms prompting is included. Universal-Streaming English and Multilingual cost $0.15/hr. Streaming is billed on how long the WebSocket stays open, so closing connections at the end of each call keeps costs down.

Is there a free API for speech-to-text?

Most providers offer free credits or a free tier rather than a permanently free API. AssemblyAI's free tier includes up to 185 hours of pre-recorded transcription and up to 333 hours of streaming, with no credit card required. Open-source models are free to download, but you pay for the GPUs and engineering time to run them.

Which speech-to-text API has the most transparent pricing?

Look for a public per-hour rate, a published price for every add-on, and clear billing rules for streaming and rounding. AssemblyAI publishes all three on its pricing page, including session-based streaming billing and per-channel billing for multichannel audio. Providers that hide rates behind a sales call or change promotional prices often are harder to forecast.

What does base speech-to-text pricing include?

Base pricing covers turning audio into text, usually with punctuation and capitalization. Features like speaker diarization, PII redaction, summarization and domain modes like Medical Mode are usually extra on unbundled providers. Some bundled providers include diarization in the base rate, so compare totals, not base rates.

How do I estimate my monthly speech-to-text costs?

Multiply your expected audio hours by the per-hour rate, add the per-hour cost of every add-on you'll use, and multiply by the channel count for multichannel audio. For streaming, use session hours, not audio hours. Then run a sample of your own audio through a free tier to confirm the accuracy matches what you need.