Cloud transcription API pricing, per hour of audio

Cloud transcription API pricing for batch speech-to-text runs roughly $0.10 to $0.61 per hour of audio at published list prices with no volume commitment, which is about one to two cents a minute at the cheap end. Real-time streaming of the same audio usually costs more, sometimes five times more, and the add-ons and the engineering around the API often cost more than the transcription itself. This page normalizes every price to one unit, dollars per hour of audio, shows the arithmetic for three common workloads, and lists what you are agreeing to when the audio leaves your machine.

How speech-to-text APIs bill

The unit is duration of audio, not file size and not words. Everything else is a modifier.

• Per second or per minute, rounded up. Google Cloud rounds each request up to the nearest one-second increment. Amazon Transcribe bills in one-second increments with no minimum. Rev AI rounds up to the nearest second but applies a 15-second minimum per job, so a two-second voice note is billed as 15 seconds. If your workload is thousands of short clips, that floor matters more than the headline rate.
• Batch and streaming are different products. Asynchronous or pre-recorded transcription is the cheap tier. Real-time streaming costs more, and at least one provider, AssemblyAI, bills streaming by how long the connection stays open rather than by how much audio you send, so an idle open microphone still meters.
• Channels multiply. Google bills each audio channel separately, so a two-channel call recording is billed as twice its length.
• Add-ons stack. Diarization, word timestamps, language identification, custom vocabulary, PII redaction, summaries and translation are priced on top. From the vendors' own pages: AssemblyAI lists diarization at $0.02 per hour on async and $0.12 per hour on streaming; Speechmatics lists translation at $0.65 per hour and summaries at $0.12 per hour; Deepgram and Amazon Transcribe both price redaction per minute on top of transcription.
• Commitments and volume tiers. Speechmatics applies a 20 percent discount above 500 hours a month. Deepgram publishes a lower Growth tier beside pay-as-you-go. Gladia publishes async rates as low as $0.20 per hour on a committed plan against $0.61 pay-as-you-go. Google, Amazon and Microsoft all publish tiered or committed pricing above the entry rate.
• Free credits run out. Deepgram offers $200 in credit, Speechmatics $100, AssemblyAI $50, Gladia 50 euros, Amazon Transcribe 60 minutes a month for 12 months, and Azure a free tier of 5 audio hours a month. These are for evaluation, not a plan.

List prices per hour of audio, batch tier

English, the entry paid tier, no volume commitment, one representative model per provider. Per-minute prices have been multiplied by 60; nothing else has been adjusted.

Provider	Model or tier priced	USD per hour of audio
Rev AI	Reverb turbo, English	$0.10
Speechmatics	Batch Melia 1	$0.129
AssemblyAI	Universal-2, async	$0.15
Microsoft Azure AI Speech	Batch transcription, standard (S1)	$0.18
Google Cloud Speech-to-Text v2	Standard model, dynamic batch	$0.18
OpenAI	gpt-4o-mini-transcribe ($0.003 per minute)	$0.18
Deepgram	Nova-3 pre-recorded, monolingual ($0.0043 per minute)	$0.26
Amazon Transcribe	Batch, US East (N. Virginia) ($0.006 per minute)	$0.36
Gladia	Async, Starter plan	$0.61

Pricing pages, in the same order: Rev AI, Speechmatics, AssemblyAI, Azure AI Speech, Google Cloud Speech-to-Text, OpenAI, Deepgram, Amazon Transcribe, Gladia.

Most providers sell more than one model, and the larger one is what you will probably end up using. On the same pages: Rev AI lists Reverb at $0.20 per hour and Whisper Large at $0.005 per minute, which is $0.30 per hour; Speechmatics lists Batch Standard at $0.24 and Batch Enhanced at $0.40; AssemblyAI lists Universal-3.5 Pro at $0.21; OpenAI lists whisper-1 and gpt-4o-transcribe at $0.006 per minute, which is $0.36 per hour; Google lists standard recognition, the non-batch path, at $0.96 per hour. Treat the table as the floor.

What streaming costs against batch

Dictation is a streaming workload, so this is the tier that applies if you are building a hotkey rather than a file pipeline.

Provider	Batch, per hour	Real-time or streaming, per hour
AssemblyAI	$0.15	$0.15 English streaming, $0.45 Universal-3.5 Pro realtime, billed on session duration
Speechmatics	$0.24 standard	$0.24 standard, $0.43 enhanced
Deepgram	$0.26	$0.29 Nova-3, listed as a promotional rate
Amazon Transcribe	$0.36	$0.60
Gladia	$0.61	$0.75
Google Cloud	$0.18 dynamic batch	$0.96 standard recognition
Microsoft Azure AI Speech	$0.18	$1.00

Two charge the same for both, three charge a modest premium, and the two hyperscalers charge roughly five times more for real time than for their cheapest batch path. OpenAI's realtime audio is priced per token rather than per minute, so it does not convert cleanly and is left out.

Three worked examples

One person dictating

30 minutes a day, 5 days a week, 52 weeks: 30 x 5 x 52 = 7,800 minutes, or 130 hours a year.

• At $0.15 per hour of streaming: 130 x 0.15 = $19.50 a year.
• At $1.00 per hour of streaming: 130 x 1.00 = $130 a year.
• Recording first and sending files instead, at $0.18 batch: 130 x 0.18 = $23.40 a year.

The trap is session billing. If your app holds the socket open two hours a day to catch half an hour of speech, you are billed for 2 x 5 x 52 = 520 hours, and at $0.15 that is $78 rather than $19.50. Per-seat dictation tools are commonly advertised in the range of $10 to $30 per user per month, which is $120 to $360 a year, so the raw API cost for one dictating user sits well under a subscription. You pay the difference in engineering.

A team recording meetings

10 one-hour meetings a week, 52 weeks: 10 x 52 = 520 hours a year.

• At $0.15 per hour: $78 a year.
• At $0.61 per hour: $317.20 a year.
• Add diarization at $0.02 per hour, which meeting transcripts need: 520 x 0.17 = $88.40 a year at the low end.

Meeting-notes products are usually sold per seat in that same $10 to $30 per user per month range. Ten seats is $1,200 to $3,600 a year against $78 to $317 of API spend, and that gap is the product: recording, joining calls, storage, search, sharing, permissions and support. The API returns a transcript, not a service. Building the meeting pipeline yourself is the fair comparison.

A 500-hour back catalog, once

• At $0.10 per hour: $50.
• At $0.15 per hour: $75.
• At $0.36 per hour: $180.
• At $0.61 per hour: $305.

A one-off archive is where cloud pricing is unambiguously good. Five hundred hours costs less than a day of an engineer's time at every price in the table, it finishes overnight because the provider runs it in parallel, and you never touch the infrastructure again.

The costs that are not in the per-hour price

• Storage and egress. If you keep the audio, you pay object storage every month and, on some providers, egress to pull it back out. Recordings dwarf transcripts in size.
• Building the integration. Auth, uploads, webhooks or polling, format conversion, chunking long files, and mapping the response into your schema. Days, not hours.
• Errors and retries. Rate limits, timeouts on long files, dropped streaming connections, and jobs that fail after you were billed for the attempt. Idempotency and a retry policy are not optional at volume.
• Monitoring. Someone has to notice when the queue backs up, when accuracy shifts after a model version changes, and when a model is deprecated.
• Key rotation and access control. API keys leak. Rotating them across environments and keeping audio buckets private is ongoing work.
• Ownership. All of it needs a name attached. A pipeline nobody owns is an outage waiting for a quiet week.

What you are agreeing to

What follows is a summary of what each vendor documents on its own pages, not legal advice and not a compliance verdict. Terms change, so confirm the current version and involve your own advisor for anything regulated. If you work under professional confidentiality obligations, private medical and legal dictation covers the questions those reviews raise.

• Google Cloud documents that Speech-to-Text does not log customer audio or transcripts by default, and that opting in to data logging, which lets Google use the data to improve the service, comes with discounted pricing. Its v1 table shows the discount plainly: $0.016 per minute with data logging against $0.024 without. See the data logging documentation.
• AWS documents that AI services may use and store customer content for service improvement, possibly in a region other than the one you are calling, and that you opt out through an AI services opt-out policy in AWS Organizations. Opting out also deletes content previously stored for that purpose.
• Microsoft documents that real-time speech to text and fast transcription are processed in server memory with nothing stored at rest, that batch transcription reads and writes storage you specify, with a timeToLive parameter to expire generated transcripts, and that diarization is off by default. See data, privacy and security for speech to text.
• OpenAI documents that API data is not used to train its models unless you opt in, that abuse-monitoring logs are retained for up to 30 days, and that zero data retention is available for eligible endpoints, including the audio transcription endpoint, subject to approval. See API data controls.
• AssemblyAI documents US and European processing, with an EU processing center in Dublin selectable from the customer dashboard, and points to its trust center for subprocessors and reports. See its security page.
• Speechmatics lists multi-region cloud options on every plan and on-premises deployment on its enterprise plan, on its pricing page.
• Deepgram's public privacy policy states that customer data is retained, stored and deleted according to its agreement with the business customer, which means the retention and training terms that apply to you are the ones in your contract, not on the website. Ask for them in writing.

The practical questions are the same for every vendor: where is it processed, how long is it kept, is it used for training by default, can that be turned off, who are the subprocessors, and can you pin a region. Read the answers before the audio goes up, not after. The wider trade-off is covered in on-device versus cloud transcription.

When the cloud API is the right answer

Plainly, and without hedging: often.

• A language or accent that local models handle poorly, where the largest hosted models are still measurably more accurate.
• A large parallel batch with a deadline. A provider runs a thousand hours across many machines at once. Your laptop runs one file at a time.
• Server-side pipelines where no user device is involved at all, such as transcribing call recordings that already live in your own storage.
• Teams that want one vendor, one support contact and one invoice rather than a fleet of machines to patch.
• Hardware that is too old or too small to run a modern model, where a per-minute bill is cheaper than new laptops.

The self-hosted middle ground

Between per-minute billing and your own laptop sits a model you run on hardware you rent, covered in self-host a transcription server, or on a single machine you already own, covered in run Whisper locally on a Mac. The catch is that a rented GPU bills by the hour whether or not audio is flowing.

A rough estimate, using a 48 GB A40 listed from $0.44 per hour on Runpod's pricing page: left running, that is 0.44 x 24 x 30 = $316.80 a month. At $0.15 per audio hour the same money buys about 2,100 hours of API transcription, and at $0.36 about 880. So an always-on box, which is what a live streaming service needs, only starts to win above roughly 900 to 2,100 hours of audio a month, and only if you keep it busy. Start it per job and stop it, and at a realistic 20 times real-time throughput that $0.44 covers about 20 audio hours, near $0.02 per audio hour of raw compute, below every API price here. Both are estimates: throughput swings with model size, batching and audio quality, and neither includes your time.

Where ThinkScribe fits

If you want a cloud dictation hotkey on a Mac instead, that build is written up here.

Frequently asked questions

How much does speech-to-text cost per hour of audio?

At list prices verified on 30 August 2026, batch transcription from the major APIs runs from about $0.10 to $0.61 per hour of audio at the entry paid tier with no commitment, with most providers clustered between $0.15 and $0.36. Larger or premium models sit higher, up to about $0.96 an hour on a hyperscaler's non-batch path, and add-ons such as diarization or redaction are charged on top.

Why is streaming transcription more expensive than batch?

Batch work can be queued and scheduled when the provider has spare capacity, so it is cheap to serve. Streaming holds a connection and a slice of GPU open for the whole session and has to return partial results within a few hundred milliseconds, which cannot be batched or deferred. Some providers also meter streaming by connection time rather than audio sent, so silence is billable.

Is a cloud API cheaper than a per-seat transcription subscription?

On raw transcription, usually yes. Ten one-hour meetings a week is 520 hours a year, which is $78 to $317 of API spend at current list prices, against $1,200 to $3,600 a year for ten seats of a per-seat product in the common $10 to $30 per user per month range. The difference pays for the product around the transcript: recording, storage, search, sharing and support. If you have to build and run that yourself, the API price is not the whole cost.

When does self-hosting a speech model beat paying per minute?

If you rent a GPU and leave it running, the crossover is roughly 900 to 2,100 audio hours a month at today's API prices, based on a 48 GB GPU listed from $0.44 an hour. If you can start the instance per job and shut it down, the raw compute cost per audio hour falls to a couple of cents and beats every API rate, but you take on the deployment, monitoring and failure handling that the API price was covering. Below a few hundred hours a month, paying per minute is almost always the cheaper and calmer choice.