Speech-to-Text API Pricing 2026: Real Cost per Hour Compared

Published on August 11, 2026 by Steven Jones

Speech-to-text API pricing in 2026 starts at roughly $0.13-$0.18 per audio hour for some managed transcription routes, but that headline number is a poor basis for choosing an API. The useful number is the cost of producing a transcript that is accurate, structured and clean enough for your application to accept.

This comparison normalises current public pay-as-you-go prices to US dollars per audio hour, then examines the costs that usually disappear from provider comparison tables: channel multiplication, speaker diarisation, redaction, real-time processing, minimum billing units, retries, storage, and human correction. Prices were checked on 11 August 2026. Enterprise contracts, temporary credits and taxes are excluded unless stated.

Speech-to-text API pricing: the quick comparison

Workload	Price-first option to test	Published API cost	Main pricing trap
Non-urgent recorded audio	Speechmatics Pro, AssemblyAI Universal-2, Google Dynamic Batch	From about $0.129 to $0.18/hour	A cheaper transcript can cost more to correct
Low-cost OpenAI transcription	gpt-4o-mini-transcribe	About $0.18/hour	Do not price new projects from the older Whisper rate alone
Real-time captions	AssemblyAI Universal-Streaming or Deepgram Nova-3	$0.15/hour and about $0.288/hour respectively	Deepgram’s current streaming figure is promotional
Two-channel contact-centre audio	Test channel-inclusive pricing carefully	Workload-dependent	Google bills channels separately while AWS includes up to two
Transcription with diarisation	Compare the combined model plus feature rate	Can add $0.02 to $0.12/hour or more	Feature charges can be large relative to the base rate
Large archive processing	AssemblyAI Universal-2, Google Dynamic Batch or OpenAI mini	About $0.15 to $0.18/hour before other costs	Do not pay real-time or standard-priority rates for work that can wait

There is no universal cheapest speech-to-text API because these routes are not equivalent. A delayed batch job, a live caption stream and a two-channel regulated customer call impose different technical requirements. Treat the table above as a shortlist for testing, not an accuracy ranking.

---

---

Current speech-to-text API cost per hour

The table below converts current provider rates into a common hourly figure. Where a provider quotes a per-minute rate, we multiply it by 60. Rates marked as promotional or “from” prices should not be used as permanent production assumptions.

Provider and route	Published price	Approx. cost per audio hour	Important condition
Speechmatics Pro	From $0.129/hour	From $0.129	Starting price rather than one universal STT rate
AssemblyAI Universal-2 async	$0.15/hour	$0.15	Recorded audio
AssemblyAI Universal-Streaming	$0.15/hour	$0.15	Real-time route
Google V2 Dynamic Batch	$0.003/minute	$0.18	Lower-priority processing
OpenAI gpt-4o-mini-transcribe	$0.003/minute estimated	$0.18	File transcription
AssemblyAI Universal-3.5 Pro async	$0.21/hour	$0.21	Higher-priced async model
OpenAI gpt-transcribe	$0.0045/minute estimated	$0.27	OpenAI’s recommended general recorded-speech route
Deepgram Nova-3 Monolingual streaming	$0.0048/minute	$0.288	Limited-time promotional streaming rate
OpenAI gpt-4o-transcribe	$0.006/minute estimated	$0.36	File transcription
Deepgram Nova-3 Monolingual pre-recorded	$0.0077/minute	$0.462	Current pay-as-you-go rate
AssemblyAI Universal-3.5 Pro Realtime	$0.45/hour	$0.45	Real-time model
Google V2 Standard	$0.016/minute at the first usage tier	$0.96	Volume tiers reduce the rate later
OpenAI gpt-live-transcribe	$0.017/minute estimated	$1.02	Live transcription
Amazon Transcribe standard Tier 1	$0.024/minute in US East pricing examples	$1.44	Region and volume affect pricing

Two details are easy to miss. Deepgram currently labels its lower streaming rates as a limited-time promotion, so it would be a mistake to build a long-term margin model around $0.0048 per minute without a fallback assumption. Speechmatics advertises Pro from $0.129 per hour and applies volume discounts above specified usage thresholds, but “from” pricing should still be validated against the exact model and processing mode you intend to use.

OpenAI pricing has also moved beyond the familiar Whisper figure. Our OpenAI Whisper API pricing guide covers the individual OpenAI routes in more detail. For a new integration, pricing only Whisper-1 and ignoring gpt-4o-mini-transcribe or gpt-transcribe can give you an outdated comparison.

Calculate the cost of an accepted transcript, not an API response

A transcription request completing successfully does not mean the output is usable. Names can be wrong, speakers can be confused, punctuation may need repair and industry terminology may need manual correction. In a subtitle workflow, timestamps can create another editing pass.

A more useful cost formula is:

Accepted transcript cost =
base recognition cost
+ channel charges
+ diarisation and redaction
+ other paid features
+ billable retries
+ storage and data transfer
+ human correction cost

The final line often deserves more attention than all the others combined.

The correction-time break-even calculation

Assume transcript review costs an illustrative $30 per working hour. Each minute of human correction then costs $0.50.

Suppose API A costs $0.15 per audio hour and API B costs $0.36. API B appears 140% more expensive. The actual difference is only $0.21 per audio hour.

Divide that $0.21 difference by the $0.50 cost of one reviewer minute, and the break-even point is 0.42 minutes, or about 25 seconds.

If the $0.36 API saves more than about 25 seconds of manual correction for every hour of audio, it is cheaper overall under this example.

This is the calculation most basic speech-to-text pricing tables miss. Tiny differences in editing time can outweigh surprisingly large percentage differences in API rates.

Channel billing can completely reverse the cheapest provider

Contact-centre recordings often store the agent and caller on separate channels. You cannot assume that one hour of stereo audio creates one billable audio hour.

Google explicitly bills each channel separately. A one-hour two-channel recording therefore represents two billable hours. Amazon Transcribe takes the opposite approach to standard transcription: its pricing includes up to two channels, so a one-hour stereo conversation is billed as one hour.

Consider 1,000 hours of two-channel calls at the first standard pricing tier:

Provider	Headline hourly rate	Billable channel-hours	Illustrative recognition cost
Google V2 Standard	$0.96	2,000	$1,920
Amazon Transcribe Tier 1	$1.44	1,000	$1,440

Amazon’s headline rate is 50% higher, yet it is $480 cheaper in this specific comparison because of the channel rule. This is not an argument that AWS is generally cheaper. It shows why pricing a contact-centre workload from a single-channel rate is unreliable.

Our dedicated Google Cloud Speech-to-Text pricing guide covers Google’s model, version and channel billing in more depth.

Diarisation and redaction should be priced before you benchmark accuracy

Speaker labels and PII redaction are not minor extras if the application requires them on every transcript. Their price should be included in the test configuration from day one.

AssemblyAI currently lists async speaker diarisation at $0.02 per hour. On Universal-2, the API cost increases from $0.15 to $0.17 per audio hour. Its real-time diarisation is materially more expensive at $0.12 per hour.

Deepgram lists speaker diarisation and redaction at $0.002 per minute each, totalling $0.12 per audio hour. Apply both to Nova-3’s current $0.0048 per minute promotional streaming rate, and the arithmetic changes like this:

Nova-3 promotional streaming recognition	$0.288/hour
Redaction	+$0.12/hour
Speaker diarisation	+$0.12/hour
Combined API price	$0.528/hour

The feature stack increases API spend by roughly 83% relative to the promotional base recognition rate. That may still be good value, but “$0.0048 per minute” would be the wrong number to put into the business model.

Minimum billing units punish short audio disproportionately

Per-second billing sounds precise until a provider also imposes a minimum charge per request. Amazon Transcribe bills in one-second increments but has a 15-second minimum per request.

Imagine a product transcribing 10,000 four-second voice clips. The actual audio duration is about 11.1 hours. At $1.44 per hour, a naive duration calculation suggests roughly $16 of transcription.

The 15-second minimum changes the billable duration to about 41.7 hours. The API cost becomes approximately $60.

Nothing about the model price changed. Request shape alone made the bill 3.75 times the simple duration estimate.

If your workload consists of voicemail fragments, voice commands, or thousands of short utterances, calculate the cost as the provider’s minimum billable unit multiplied by the request count. Do not calculate from total WAV or MP3 duration alone.

Real-time transcription needs a different cost scorecard

Batch and streaming prices should not be combined into a single ranking. A batch archive can tolerate a queue. Captions, voice agents, and live call assistance cannot.

For 1,000 hours of live audio, representative current base prices produce the following API spend:

Real-time route	Approx. cost/hour	1,000 audio hours
AssemblyAI Universal-Streaming	$0.15	$150
Deepgram Nova-3 Monolingual streaming promotional rate	$0.288	$288
AssemblyAI Universal-3.5 Pro Realtime	$0.45	$450
Google V2 Standard at first tier	$0.96	$960
OpenAI gpt-live-transcribe	$1.02	$1,020
Amazon Transcribe Tier 1	$1.44	$1,440

This is a cost table, not a real-time quality ranking. For a production streaming application, test time to first partial result, finalisation behaviour, word confidence, handling of interruptions, reconnect behaviour and accuracy on the actual microphones and accents you expect.

A recurring issue among developers building speech applications is that price ceases to be the deciding factor once the live behaviour is poor enough to require workarounds. A cheap recogniser that finalises at awkward points, lacks information your application needs, or forces additional buffering can create more engineering work than its API savings justify.

Six real speech-to-text pricing scenarios

1. Transcribing 100 podcast hours

For mono podcast recordings that can be processed asynchronously, API-only spend is remarkably small:

Route	100 audio hours
Speechmatics Pro starting rate	From $12.90
AssemblyAI Universal-2	$15
Google Dynamic Batch	$18
OpenAI gpt-4o-mini-transcribe	About $18
OpenAI gpt-transcribe	About $27

Now add 5 minutes of human correction per audio hour at our illustrative $ 30-per-hour reviewer rate. Review costs $250 across the 100-hour archive. The difference between a $15 and a $27 API has become almost irrelevant compared to the editing bill.

For podcast transcription, therefore, test names, punctuation, speaker separation and specialist vocabulary before chasing a few cents of API savings.

2. Processing 1,000 customer-call hours with PII redaction

Feature charges become much more visible at contact-centre scale. Using Deepgram’s current Nova-3 Monolingual promotional streaming rate as a simple example, 1,000 audio hours cost about $288 for recognition. Adding its $0.12-per-hour redaction feature raises that to roughly $408.

Add speaker diarisation as well, and the API component reaches about $528. Procurement should compare that $528 configuration with equivalent configurations elsewhere, not compare the $288 base figure with another provider’s all-in rate.

3. Running 1,000 hours of real-time captions

The base-price spread in our current table runs from $150 for 1,000 hours of AssemblyAI Universal-Streaming to $1,440 at Amazon Transcribe’s first-tier $0.024-per-minute rate. That looks decisive until latency, language support, finalisation, timestamps, confidence data and transcript quality are tested.

For captions, a delayed or unstable result has a product cost even if it has no line on the transcription invoice. Build the evaluation around live sessions rather than feeding prerecorded files to a streaming endpoint and assuming the result represents production behaviour.

4. Transcribing 1,000 hours of dual-channel contact-centre audio

This is where billing architecture can outweigh the headline rate. At the first standard tier, Google’s $0.96-per-hour equivalent becomes $1.92 per original file hour when two channels are billed separately. Across 1,000 stereo hours, that produces the $1,920 recognition estimate shown earlier.

Amazon’s $1.44-per-hour Tier 1 figure looks dearer until its included two-channel treatment is applied, resulting in $1,440 for the same nominal 1,000-hour workload. The next comparison should be output quality and required features, not the original cents-per-minute figures.

5. Transcribing 1,000 recorded hours with diarisation

AssemblyAI Universal-2, plus its $0.02-per-hour async speaker diarisation, yields an API price of about $0.17 per audio hour, or $170 for 1,000 hours. Universal-3.5 Pro plus the same async diarisation charge is about $0.23 per hour, or $230.

The $60 difference only buys value if the higher-priced configuration produces better transcripts for your audio. At a $30-per-hour review rate, it needs to save only two minutes of correction across each 1,000 audio hours in aggregate to repay $1, but the useful comparison is per audio hour: a $0.06 hourly API premium breaks even after saving roughly seven seconds of reviewer time per audio hour.

6. Processing a 10,000-hour archive

Archive transcription is where paying for urgency becomes particularly wasteful. Ten thousand hours is 600,000 minutes of audio. At current public rates, AssemblyAI Universal-2 would cost about $1,500 for base transcription, while Google Dynamic Batch and OpenAI gpt-4o-mini-transcribe are each around $1,800 before overhead costs.

Using Google V2 Standard instead of Dynamic Batch changes the calculation sharply. The first 500,000 minutes at $0.016 cost $8,000, and the next 100,000 minutes at $0.01 cost another $1,000. Total recognition cost is about $9,000.

If the archive does not need prompt turnaround, that $7,200 gap between Google Standard and Dynamic Batch is the price of choosing the wrong processing class, not the wrong speech model.

Retries, storage and data transfer belong in the model, but do not invent them

Some cost comparisons add arbitrary fixed estimates for Cloud storage, bandwidth and infrastructure. That makes the spreadsheet look complete while making the result less useful.

Use your actual architecture instead. If audio already resides in the same cloud and region as the transcription service, transfer behaviour can look very different from that of a pipeline that moves recordings between providers. Long-term archive storage also has a different cost profile than temporary staging objects that are deleted after transcription.

Google’s official Speech-to-Text pricing documentation, for example, explicitly separates Speech-to-Text charges from those for additional Google Cloud resources, such as storage.

The same discipline applies to retries. Measure the percentage of audio that is actually reprocessed and billable in your implementation. Do not add a generic “5% retry allowance” simply because a comparison article tells you to.

Do not compare self-hosted Whisper with API prices using GPU rental alone

Self-hosted speech recognition can become economically attractive at high, predictable utilisation. Comparing a GPU’s hourly rental price with an API’s audio-hour price is still incomplete.

A production self-hosted calculation also needs GPU utilisation, idle capacity, batching efficiency, model loading, queues, autoscaling, deployment time, monitoring, failure recovery, storage, networking and engineering ownership. If a GPU is idle for most of the day, impressive peak transcription throughput does not translate directly into impressive unit economics.

Practitioners comparing local and hosted speech recognition repeatedly run into this operational boundary. Batch workloads with sufficient predictable volume can favour local models, while SaaS and real-time applications often value managed scaling and easier provider switching over theoretical compute efficiency.

A sensible self-hosting test therefore calculates cost per accepted audio hour at actual utilisation, not cost per GPU hour at 100% utilisation.

How to benchmark speech-to-text pricing properly

A short evaluation is enough to reject bad candidates, but only if you use representative audio. A polished ten-minute recording is useless for predicting what happens to noisy telephone calls.

• Build a representative audio set. Include the microphones, accents, languages, background noise, crosstalk and specialist terminology found in production.
• Run the exact production mode. Test streaming as streaming, dual-channel audio as dual-channel audio and every required paid feature enabled.
• Record billable duration. Keep both the actual audio duration and the provider-billed duration so that rounding and minimum charges are visible.
• Measure correction time. Track human minutes required to reach your acceptance standard rather than relying on word error rate alone.
• Record failure and retry behaviour. Include rejected files, reconnections, timeouts and any duplicate processing that becomes billable.
• Calculate accepted-transcript cost. Combine API spend, features and review labour for each provider.
• Test at expected scale. Apply the correct volume tier and concurrency requirements rather than multiplying a small free-credit experiment indefinitely.

The cheapest speech-to-text API depends on what creates the bill

If you only need low-priority transcription of clean recorded audio, current managed API entry points around $0.13 to $0.18 per audio hour make transcription remarkably inexpensive. Speechmatics Pro, AssemblyAI Universal-2, Google Dynamic Batch and OpenAI’s lower-cost transcription route all deserve consideration on price.

The ranking changes as soon as the workload changes. Real-time audio has different economics. Two-channel recordings expose different billing rules. Diarisation and redaction can substantially increase an apparently cheap rate. Very short clips can be punished by request minimums. At scale, choosing standard-priority processing for an archive can waste thousands of dollars.

Human correction is the final filter. If a more expensive API saves even a few seconds of editing per audio hour, it can beat a cheaper model in total cost. Build the spreadsheet around the transcript you can actually use, and the provider choice becomes much easier to defend.