Software advice
Best AI Speech to Text APIs in 2026
An AI speech to text API is a metered service that developers call with audio, either a finished file or a live stream, and that returns text, often with speaker labels and timestamps. It is billed by audio duration or tokens. It is not a finished transcription app with an editor, and it is not text to speech.
Quick answer
Deepgram ranks first because its streaming and batch rates, concurrency caps and a no-card credit are all printed. Soniox is the pick for the lowest published hourly rate, and DeepScript for German hosting and a stated no-training policy.
A developer choosing a speech to text API faces a price list written in four different units: per hour, per minute, per second and per million tokens. Before comparing quality, someone has to turn those into a cost per hour of audio and check how many live streams the plan allows. This ranking judges what the vendors print: rates, free allowances, concurrency limits, deployment options, and what the pages say about retention and model training. We read vendor pages in November 2026 and ran no audio through any service, so accuracy and latency claims are the vendors' own. Where a pricing page did not render its rates, the entry says so instead of guessing.
The 12 best AI speech to text APIs in 2026, ranked
There is no score. Vendors can pay for more visibility on this website, for example a higher place in a ranking or more room on a page. Commercial relationships can also affect which products we feature and how prominently. How we rank.
| # | Tool | Best for | Free plan | Entry price |
|---|---|---|---|---|
| 1 | Deepgram | Teams building live transcription or voice agents that need printed rates and concurrency numbers | Trial | Free $200 credit, then pay as you go |
| 2 | Soniox | High-volume buyers who want the lowest published hourly rate and bundled diarization | No | $0.10 per hour |
| 3 | Amazon Transcribe | Teams already on AWS that want per-second billing and a 12-month free allowance | Trial | $0.006 per minute (batch, US East example) |
| 4 | DeepScript | European teams that want German servers, a no-training statement and per-hour pricing | Trial | 0,18 euros per hour (Standard) |
| 5 | AssemblyAI | Developers who want a no-card trial and published streaming concurrency per tier | Trial | $50 in free credits on signup |
| 6 | Azure AI Speech | Microsoft-centered organizations that want real-time, batch, LLM and custom transcription tiers in one service | Yes | Quoted |
| 7 | OpenAI speech-to-text API (gpt-4o-transcribe) | Teams already using the OpenAI API that want transcription on the same account | No | $2.5 per 1M audio input tokens |
| 8 | Corti | Healthcare AI builders who need clinical conversation transcription with EU residency options | Trial | $0.0065 per minute (Speech to Text) |
| 9 | ElevenLabs | Creators already paying for ElevenLabs audio products who want transcription from the same credits | Yes | $0 per month (Free plan) |
| 10 | pyannoteAI | Teams that need speaker diarization or speaker-attributed transcripts more than plain transcription | Trial | 19 euros per month (Developer) |
| 11 | Voxist | French-speaking or privacy-first teams that want real-time transcription with a stated no-storage policy | No | Quoted |
| 12 | Scriptix | Dutch and European media or public-sector teams that want transcription, subtitles and translation in one workflow | Trial | Quoted |
- 1
Deepgram ranks first because a buyer can price a deployment from the page alone. It prints streaming rates for monolingual and multilingual Nova-3, a cheaper Growth tier, concurrency caps and a $200 credit that does not expire. Overage and enterprise terms are visible too. The gaps are hosting region, retention and training policy, none of which the pricing page states.
Strong on
- Prints streaming and pre-recorded rates, so cost per audio hour can be calculated directly
- Free $200 credit needs no card and does not expire
- Enterprise adds custom model training and a self-hosted deployment option
Weak on
- Growth overage bills at the Growth rate plus 10 percent, charged weekly
- The pricing page states no retention, training or hosting region policy
Skip it if: Buyers who need a published hosting region or retention period before they can trial the service. Nova-3 streaming is $0.46 per hour on pay as you go. Growth starts at $4,000 a year.
- 2
Soniox takes second place on price clarity: $0.10 per hour for files and $0.12 for streaming, with diarization, language identification and smart formatting bundled into the real-time rate. It loses a place to Deepgram because no free tier is mentioned, the legal entity and country are unstated, and data handling was not covered by the pricing page.
Strong on
- Lowest published hourly rates among the developer platforms we read
- Diarization and language identification are bundled into the real-time rate
- Hourly equivalents are printed next to the token prices
Weak on
- No free tier is mentioned on the pricing page
- Billing is token-based, so hourly figures are the page's own equivalents
Skip it if: Teams that want a no-cost trial credit or written data-handling terms on the pricing page. File transcription is $0.10 per hour and real-time is $0.12. No free tier is mentioned.
- 3
Amazon Transcribe ranks third because per-second billing with no minimum and separate batch and streaming rates make costs predictable, and the 60 free minutes a month last a year. The tiered regional tables did not render for us, so only example rates are known, and the free tier excludes content redaction and custom language models.
Strong on
- Per-second billing with no minimum removes rounding waste on short clips
- Free tier gives 60 audio minutes a month for 12 months
- Two-channel audio is billed on total duration, not per channel
Weak on
- Regional and tiered rate tables did not render, so only example rates are verified
- Free tier excludes content redaction and custom language models
Skip it if: Buyers who need a stated retention or training policy from the pricing page alone. Streaming is $0.01 per minute in the same example. Billing is per second with no minimum.
- 4
DeepScript is the clearest on data handling in this list: servers in Germany, a homepage statement that data is never used for training, and automatic deletion after 7 to 365 days on request. Prices are low and printed. It sits fourth because the free allowance is only three transcriptions and no certifications are named on the page.
Strong on
- Servers in Germany and a stated no-training policy on the homepage
- Prices of 0,18, 0,27 and 0,29 euros per hour are printed
- Browser upload, live transcription, REST API and a Claude MCP integration
Weak on
- Automatic deletion is set on request, not by default
- No certifications are stated on the opened page
Skip it if: Teams that need named compliance certifications or published concurrency numbers. Three transcriptions are free, then usage billing with no subscription.
- 5
AssemblyAI ranks fifth. The $50 signup credit needs no card and the pricing page states streaming concurrency for both tiers, 5 new streams per minute free and 100 on pay as you go. The hourly rates were not captured in our reading and volume terms are quote-only, which makes budgeting harder than at the vendors above.
Strong on
- A no-card $50 signup credit covers early testing and prototypes
- Streaming concurrency is printed: 5 new streams per minute free, 100 on pay as you go
- Volume pricing and commit buyouts are available through sales
Weak on
- The free tier is a signup credit only, with no monthly free hours named
- Rate card figures were not captured, so hourly cost needs a second look
Skip it if: Buyers who need a full price list on one page or a published retention policy. Rates are usage-based rate cards. Streaming concurrency is 5 new streams per minute on free.
- 6
Azure AI Speech has the widest tier list: real-time, fast, batch, MAI Transcribe models, an LLM-based tier and custom models, plus commitment tiers up to 50,000 hours a month. It ranks sixth because its per-hour rates did not render, so price cannot be confirmed from the page, and custom endpoint models are removed after 7 idle days without use.
Strong on
- Free F0 tier includes 5 audio hours of real-time transcription each month
- Diarization and language identification are sold as add-ons per hour
- Commitment tiers at 2,000, 10,000 and 50,000 hours a month
Weak on
- Per-hour rates did not render on the opened page
- Custom endpoint models unused for 7 days are decommissioned
Skip it if: Buyers who need a price they can read from the page or batch transcription on the free tier. Free F0 gives 5 audio hours a month. Paid per-hour rates did not render on the page.
- 7
OpenAI's gpt-4o-transcribe ranks seventh. It is easy to add for teams already on the platform, and the model page prints the audio input price. But billing is by token with no per-minute figure, there is no stated free tier, and the page says nothing about retention or training, so the buyer has to measure and ask.
Strong on
- Transcription sits on the same account as OpenAI's other models
- Audio input price is printed at $2.5 per 1M tokens
Weak on
- No per-minute price, so cost per audio hour must be estimated
- The model page states no retention or training policy
Skip it if: Buyers who want a plain per-hour rate or a free tier for testing. The price is per token, with no per-minute figure on the model page. No free tier is stated.
- 8
Corti is a specialist: its Speech to Text API covers dictation and clinical conversations in about 50 languages at $0.0065 per minute, and Enterprise offers EU residency or on-premises hosting. It ranks eighth because the free credit amount is not published, its compliance page was not opened in our reading, and the focus on healthcare is narrow for general buyers.
Strong on
- Stated use case is clinical conversations, with streaming and batch audio
- EU residency or on-premises hosting offered on Enterprise
- Startup program with extended credits for healthcare AI founders
Weak on
- The amount of free signup credit is not published on the pricing page
- The pricing page mixes speech rates with token rates for non-speech models
Skip it if: General-purpose transcription buyers outside healthcare who want a broad language and plan range. Usage-based with no plan fee. Free signup credits, but the amount is not published.
- 9
ElevenLabs ranks ninth because transcription is a metered add-on to a credit pool built for other audio products. At 330 credits per minute the free plan covers about 30 minutes, and Pro's 600,000 credits cover roughly 30 hours. The monthly pool is simple to understand, but the pricing page states no retention, training or hosting policy.
Strong on
- Speech to Text draws on one credit pool shared with other products
- Plans run from Free to Business and Enterprise with printed credit counts
Weak on
- Free plan covers about 30 minutes of transcription a month, without rollover
- Retention, training and hosting policy are not stated on the pricing page
Skip it if: High-volume transcription buyers, since credits at 330 per minute cost more per hour than usage-billed rivals. Speech to Text costs 330 credits per minute from a shared monthly pool of 10,000 on Free.
- 10
pyannoteAI ranks tenth because it is built around who spoke, not plain transcription. Its STT Orchestration pairs diarization with Parakeet v3 0.6b or Whisper Large v3 turbo. Monthly credit plans are priced clearly in euros, but the trial hours conflict on the page and it is a narrower tool than the full transcription platforms above.
Strong on
- Streaming and batch diarization are listed on the Developer and Starter plans
- Speaker identification with voiceprints at 0.015 euros per voiceprint
- Plans are priced in euros, with usage credit equal to the monthly fee
Weak on
- Trial hours conflict on the page, 170 hours in one place and 150 in the FAQ
- Developer plan includes one seat, Starter three users
Skip it if: Buyers who only need plain transcripts and would pay for diarization they do not use. Developer includes 19 euros of usage credit per month. Starter is 99 euros with 99 euros of credit.
- 11
Voxist ranks eleventh. Its homepage makes the strongest privacy statements after DeepScript: nothing is stored, no audio trains models, and audio does not cross a border the customer did not choose. But no API price is published, the pricing link leads to a coming-soon page, and no hosting region or certification is named on the homepage.
Strong on
- Homepage states that nothing is stored and no audio is used for training
- Real-time transcription of streamed audio through the Voxist API
Weak on
- No API prices are published and the pricing page is marked coming soon
- Hosting region is not named, only that audio does not cross a border the customer did not choose
Skip it if: Buyers who must compare prices before talking to sales or who need a named hosting region. The Voxist API is priced in euros, but no amounts are published and the pricing link is a coming-soon page.
- 12
Scriptix ranks last because its homepage gives the least to compare: no prices, a trial only by request, and no stated data practice at all. What it does show is a workflow product rather than a bare endpoint: Transcribe, Subtitle and Translate tools, custom speech models and glossaries, a speaker directory, and cloud or on-premise deployment.
Strong on
- Transcribe, Subtitle and Translate workflows in one product
- Custom speech models, glossaries and a speaker directory
- Cloud and on-premise deployment are both named on the homepage
Weak on
- No prices on the opened page and the trial is by request
- Retention, training and hosting policy are not stated on the opened page
Skip it if: Developers who want to price an API from a public rate card and start in minutes. Prices are not shown on the opened page. A trial is requested through a form.
How these AI speech to text APIs were judged
How accurate are speech to text APIs, and what do the vendors actually claim?
Accuracy is the criterion this ranking can judge least, because none of the pages we read publish a word error rate for a named test set. Soniox calls its STT v6 model the world's most accurate, which is a marketing claim, not a measurement. Deepgram names Nova-3 in monolingual and multilingual versions, Azure lists MAI Transcribe models and an LLM-based tier, and OpenAI names gpt-4o-transcribe. What the pages do show is feature breadth: diarization, language identification and smart formatting at Soniox, speaker-attributed output at pyannoteAI, and about 50 languages at Corti. Test your own recordings, with your accents and background noise, before trusting any claim.
How do you compare speech to text prices when every vendor bills differently?
Convert everything to dollars per audio hour. Soniox prints $0.10 per hour for files and $0.12 for real-time. Deepgram's pay-as-you-go streaming rate for Nova-3 is $0.46 per hour, with a $0.39 rate on its Growth plan, which starts at $4,000 a year. Amazon lists $0.006 per minute for batch and $0.01 for streaming in its US East example, which works out to $0.36 and $0.60 per hour. OpenAI bills tokens only, so you estimate the hour yourself. Azure's rate table did not render when we read it. Free allowances differ sharply: Deepgram gives $200, AssemblyAI $50, Azure 5 hours monthly, and Amazon 60 minutes monthly for 12 months.
Where does your audio go, and is it used to train models?
Most vendor pricing pages say nothing about retention or training, and a buyer should treat silence as an open question to put to sales or to the privacy policy. The clear statements here come from small vendors. DeepScript says on its homepage that customer data is never used for training, offers automatic deletion after 7 to 365 days on request, and names servers in Germany. Voxist states that nothing is stored and no audio trains models. Corti offers EU residency or on-premises hosting on its Enterprise tier, and Deepgram and pyannoteAI list self-hosted or on-premise options for enterprise contracts. For the rest, we did not verify hosting regions or certifications.
How hard is it to switch speech to text providers later?
The API call itself is simple to swap, since every vendor takes audio and returns text. The harder parts are what surrounds it: custom vocabularies, custom models, speaker voiceprints, and word timestamps formatted in one vendor's schema. Azure's custom endpoints are billed per model per hour and are decommissioned after 7 days without use, and Scriptix sells custom speech models and glossaries inside its own editors. Billing units also trap budgets: a move from per-minute to per-token pricing changes your forecast. Keep a test set of 50 recordings and a script that scores any provider against it, so a switch is a one-afternoon experiment rather than a project.
Which speech to text API fits a startup, an enterprise or a regulated team?
A startup prototyping a voice feature should pick the vendor with the largest no-card credit, which currently means Deepgram's $200 or AssemblyAI's $50, and watch the streaming concurrency caps. A team already running on AWS or Azure may accept a slightly higher rate to keep billing and access control in one place. A clinic or a European public body needs written answers on hosting and retention first, which puts Corti, DeepScript and Voxist ahead of vendors whose pages are silent. A product that needs to know who spoke, such as meeting or interview software, should look at pyannoteAI's diarization-first plans.
How to choose an AI speech to text API in three steps
- Convert every price to dollars per audio hour. Open each pricing page and rewrite the rate in one unit: dollars per hour of audio, split into batch and streaming. Multiply per-minute rates by 60 and divide token prices by the tokens your audio actually produces, which you can measure on a ten minute sample. Then add the extras that are billed separately, such as diarization at Azure or voiceprints at pyannoteAI. Compare the total at your expected monthly hours, and again at ten times that volume, because several vendors change rate or plan at the volume boundary.
- Check concurrency before you check accuracy. A live product fails first at the concurrency limit, not at the word error rate. Deepgram prints 150 concurrent streams on pay as you go and 225 on Growth, while AssemblyAI prints 5 new streams per minute on its free tier and 100 on pay as you go. Write down your expected peak of simultaneous calls, then match it against each cap and ask what happens on the sixth connection. Tiers with no published cap should get a written answer from sales before launch, not after.
- Run your own recordings and read the data terms. Collect 50 clips that look like production: phone audio, accented speakers, crosstalk, domain words. Send the same clips to your three shortlisted services and score the errors that cost you money, such as names, numbers and product terms. In the same week, ask each vendor three questions in writing: is audio retained and for how long, is it used for training, and in which region is it processed. Pricing pages usually do not answer them, and the answers decide which vendors stay on the list.
Frequently asked questions
What is the best AI speech to text API in 2026?
Deepgram ranks first here because it prints streaming and batch rates, concurrency caps, a Growth plan and a $200 credit that needs no card. Soniox has the lowest published hourly rate we found, and DeepScript is the pick for servers in Germany and a stated no-training policy. Accuracy was not tested.
Which speech to text API is cheapest per hour?
Among published rates, Soniox lists $0.10 per hour for file transcription and $0.12 for real-time, and DeepScript lists 0,18 euros per hour on Standard. Amazon works out to $0.36 per hour batch from its $0.006 per minute example. Corti lists $0.0065 per minute, about $0.39 per hour. Rates change, so check each page.
Is there a free speech to text API with no credit card?
Deepgram gives a $200 credit with no card and no expiry, and AssemblyAI gives $50 on signup without a card. Azure's F0 tier includes 5 audio hours of real-time transcription each month, and Amazon gives 60 minutes monthly for 12 months. ElevenLabs Free has 10,000 credits, roughly 30 minutes of speech to text.
Which speech to text APIs offer EU hosting or on-premises deployment?
DeepScript names servers in Germany. Corti offers EU residency or on-premises hosting on its Enterprise tier, and pyannoteAI offers on-premise on Enterprise. Deepgram lists self-hosted options for Enterprise, and Scriptix names cloud and on-premise deployment. Voxist says audio never crosses a border the customer did not choose. Verify the region in a contract before relying on it.
Can a speech to text API tell speakers apart?
Yes, several do. Soniox bundles diarization into its real-time rate, Azure sells diarization as a per-hour add-on, and pyannoteAI is built around it, with streaming and batch diarization and speaker identification at 0.015 euros per voiceprint. Scriptix lists a speaker directory. Check whether diarization works on your live streams as well as on files.
How do per-token prices from OpenAI compare with per-minute prices?
OpenAI lists gpt-4o-transcribe audio input at $2.5 per 1M tokens and gives no per-minute figure, so a buyer must measure how many tokens an hour of audio produces. Soniox also bills tokens but prints hourly equivalents. Run a ten minute sample, read the token count from the response, and scale it to an hour.
What happens when I exceed the limits of a speech to text API plan?
It depends on the vendor. Deepgram's Growth plan bills overage at the Growth rate plus 10 percent, charged weekly. AssemblyAI caps new streams per minute by tier. ElevenLabs draws from one credit pool that does not roll over on Free. Azure's free tier stops at 5 hours monthly. Read the overage clause before you rely on a free or entry plan.
Do speech to text vendors train their models on my audio?
Few pricing pages say. DeepScript states customer data is never used that way, and Voxist says no audio is used for model training. Deepgram, AssemblyAI, Amazon, Azure, OpenAI, Soniox, ElevenLabs and Corti did not state a training policy on the pages we read, so ask for it in writing or read the privacy policy before sending sensitive recordings.