Speech-to-text APIs turn call recordings, meetings and live voice streams into text that your own software can search and act on.
This guide ranks twelve services for developers and buyers, weighing the real hourly bill after add-ons, how audio can be deployed, and whether a small vendor can be adopted without a contract.
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. Every vendor link is marked the same, whatever the arrangement behind it. How that works.
In short
What speech-to-text apis software does
A speech-to-text API receives audio files or live streams over a programming interface and returns transcripts with timestamps, speaker labels and optional analysis, billed by audio duration or credits.
Our order comes from five tests, taken in turn: effort to set up, what you really pay, what happens to your data on the way out, how locked in you are, and the intended customer. One rule overrides the order: smaller vendors take the first three places when the category offers them, and the best-known brands start at four. We explain on how we work why the tests are these five and why no product receives a score.
Hourly speech API with speaker detection and translation bundled in
Ranked #1 of 12 in Best Speech-to-Text APIs in 2026.
Published pricingElsewhere
The pitch is one hourly price with the extras already inside it. Diarization, language detection and formatting are all bundled, and translation rides along in the real-time call. Across 60+ languages it copes with speakers who switch language mid-sentence.
Regional processing exists for data residency, and Perplexity is named as an integrator. Against that, no trial is described, nothing says the software can run on your own servers, and the terms page returned not found when checked.
What stands out
Bundled diarization
Mid-sentence language switching
Regional processing
Where it costs you
Testing means paying from the first hour, since nothing on the page offers a sample allowance.
Only a cloud API is described, with regional processing as the nearest thing to local control.
Right for
Teams transcribing multilingual calls who want one hourly rate with the extras included
Wrong for
Buyers who must run recognition inside their own data centre
Not stated by the vendorPublished per audio hour, with asynchronous and real-time rates; enterprise discounts on request
Per-hour transcription with starter credit and a long add-on menu
Ranked #2 of 12 in Best Speech-to-Text APIs in 2026.
Published pricingElsewhere
Two model families appear on the pricing page: Universal-2 and Universal-3.5 Pro for recorded audio, with separate streaming and Realtime options at higher rates. Beyond transcription sit PII redaction, content moderation, an LLM Gateway billed per million tokens, and reference builds for LiveKit and Pipecat voice agents.
Compliance lists HIPAA BAA, ISO 27001, SOC 2 Type 2 and GDPR, with EU data residency at matching prices. Everything shown is vendor-hosted; the pages offer no way to run it yourself.
What stands out
Add-on pricing
Signup credit
EU data residency
Where it costs you
Feature-heavy jobs cost more than the headline figure because each extra is its own hourly line.
The starter allowance limits streaming to 5 new streams per minute, against 100 on pay-as-you-go.
Right for
Developers who want to start on credits and pick analysis features one by one
Wrong for
Teams that need one flat all-in hourly price
Not stated by the vendorPublished per audio hour by model; starter credit at signup; enterprise volume deals by contract
Model-tiered hourly pricing with deployment on device, on prem or cloud
Ranked #3 of 12 in Best Speech-to-Text APIs in 2026.
Self-hostablePublished pricingElsewhere
Pricing is split by model: Melia 1 for pre-recorded and real-time, Standard, Enhanced, Oak 1 for medical dictation and Linden 1 for voice agents. Self-serve buyers choose pay as you go, subscriptions or credit packs, and the Free tier allows 2 concurrent real-time sessions against 50 on self-serve.
SOC 2 Type II, ISO/IEC 27001:2022, GDPR and HIPAA are listed, and processing can sit in the EU, US or Australia. Vapi, Pipecat, LiveKit and Zapier are named integrations.
What stands out
Model tiers
Medical model
On-prem containers
Where it costs you
Anything other than the vendor cloud needs an Enterprise contract.
Single sign-on costs extra every month, and there is no headquarters on the home page.
Right for
Teams wanting model choice by price tier plus a route to on-device deployment
Wrong for
Small projects that need containers without an Enterprise contract
Not stated by the vendorPublished per hour by model tier, with subscriptions, credit packs and Enterprise contracts
Per-minute streaming and recorded rates, prepaid credits, containers for Enterprise
Ranked #4 of 12 in Best Speech-to-Text APIs in 2026.
Self-hostablePublished pricingElsewhere
Pay As You Go allows 150 concurrent streaming connections at Deepgram and Growth allows 225, and Voice Agent concurrency tops out at 60 in North America on Growth. Growth also trims the Nova-3 minute rates by about 16%.
Smart Formatting costs nothing extra, and diarization, redaction, entity detection, summarisation, topic detection, sentiment and intent recognition are listed. Compliance includes SOC 2 Type 2, HIPAA with a BAA on Enterprise, GDPR with an EU endpoint and FIPS 140-3 images for containers.
What stands out
Nova-3 models
Prepaid Growth credits
Self-hosted containers
Where it costs you
Self-hosted containers need NVIDIA GPUs and an Enterprise contract.
Overages on Growth are billed at the Growth rate plus 10%, which blunts the prepaid discount.
Right for
Teams with steady streaming volume ready to prepay for a lower minute rate
Wrong for
Buyers wanting an on-premise deployment without an Enterprise contract
Not stated by the vendorPublished per minute with no minimum; prepaid annual Growth credits; Enterprise by contract
Apache-licensed offline speech toolkit that runs on phones and Raspberry Pi
Ranked #5 of 12 in Best Speech-to-Text APIs in 2026.
Free tierOpen sourcePublished pricingElsewhere
Small models of about 50 MB per language make it practical on a Raspberry Pi or a phone, and a Vosk Server handles larger server models. Bindings exist for Python through pip, Java, C# and JavaScript on Node, with a streaming API that allows vocabulary reconfiguration and speaker identification.
English, German, French, Spanish, Chinese, Russian, Hindi and Korean are among the supported languages. The page gives no accuracy figures, so measure it on your own audio.
What stands out
Apache-2.0 licence
Offline recognition
Python, Java and C# bindings
Where it costs you
Nobody sells you a service level here, as the vendor page shows only the toolkit.
Language coverage is 20+ languages, narrower than the 57+ and 60+ that cloud rivals list.
Right for
Developers embedding offline recognition in devices, kiosks or air-gapped systems
Wrong for
Teams wanting a managed API with a service agreement
Not stated by the vendorFree, Apache-2.0 open source code; the vendor page lists no cloud service or prices
Usage-based transcription with a human option and on-prem deployment
Ranked #6 of 12 in Best Speech-to-Text APIs in 2026.
Self-hostablePublished pricingElsewhere
The Pay As You Go page prices Reverb English at one rate, Reverb foreign language higher, Whisper Large per minute, forced alignment per minute and human transcription per minute, rounded up to the second.
Capabilities include language identification, sentiment analysis, topic extraction, summarisation and translation across 57+ languages. SOC II, HIPAA, GDPR and PCI are listed, with a 99.99% uptime claim. Enterprise adds a dedicated account manager, priority technical support and extra evaluation credits.
What stands out
Human transcription option
Forced alignment
Cloud or on-prem
Where it costs you
The page names no third-party integrations, only SDKs and a developer API.
Rev AI's own pages never say where the company is based.
Right for
Teams wanting both machine and human transcripts from a single vendor account
Wrong for
Buyers who need named integrations listed on the vendor pages
Not stated by the vendorPublished per hour or minute by product; starter credit of 5 hours; Enterprise by conversation
Paris-based voice platform offering ASR with air-gapped deployment choices
Ranked #7 of 12 in Best Speech-to-Text APIs in 2026.
Pricing on requestEurope
Beyond the API, Voxist lists Voxcept, Voxlive for contact centres, Voxlingo for real-time voice translation, Voxreply for voicemail and Voxmemo for dictation. Named partners include OVHcloud, Scaleway, Cisco, Deutsche Telekom and Northern Data Group, and an HDS-hosted option exists.
GDPR-native and EU AI Act readiness are claimed. Sub-200 ms first audio and 40+ languages for Voxlingo are the vendor's own figures, so treat them as claims to test, not results.
What stands out
Air-gapped option
French recognition
HDS-hosted
Where it costs you
Pricing is a placeholder page, which rules out any like-for-like cost comparison.
Certification is unfinished: SecNumCloud sits on the roadmap.
Right for
French-language buyers who need hosting that stays inside a controlled perimeter
Wrong for
Anyone who must compare per-hour prices before a sales call
FrancePriced in EUR, though the pricing link leads to a coming-soon page and shows no amounts
Credit-pool plans where speech-to-text draws on a shared monthly balance
Ranked #8 of 12 in Best Speech-to-Text APIs in 2026.
Free tierPublished pricingElsewhere
Plans run Free, Starter, Creator, Pro, Scale, Business and Enterprise, with credit allowances rising from 10k to 6M a month. Because every product draws on that pool, the cost of transcribing an hour depends on what else the account spends credits on.
At 330 credits per minute, an hour consumes 19,800 credits, which is about a sixth of the Creator allowance. No Scribe-specific features were listed on the pages read, so capability claims need checking in the documentation.
What stands out
Shared credit pool
Free plan included
Metered by the minute
Where it costs you
No per-hour speech-to-text rate is shown, and Scribe is not named on the pricing page.
Credits are shared with other products, so transcription competes with them for the same balance.
Right for
Buyers already on ElevenLabs credits who want transcription from the same account
Wrong for
Teams needing a stated per-hour rate for budgeting
Not stated by the vendorMonthly subscriptions with a free plan and a shared credit pool; Enterprise priced by contract
MIT-licensed speech model you run yourself, or call through a per-minute API
Ranked #9 of 12 in Best Speech-to-Text APIs in 2026.
Published pricingElsewhere
Two routes exist. Locally, the MIT-licensed model loads through pip and a load_model call, which suits batch jobs where audio must stay on your machine.
Through the API, Whisper and gpt-4o-transcribe share one minute rate, and the latter also lists token prices of $2.50 input and $10.00 output per million, and gpt-4o-mini-transcribe is $0.003 per minute. The pricing page lists transcription only, with no diarization or analytics line items.
What stands out
MIT licence
Per-minute API
Token-priced options
Where it costs you
No credits or trial are listed for the transcription models.
Running it locally means you supply the hardware, scaling and monitoring.
Right for
Developers who want a permissively licensed model with a hosted fallback
Wrong for
Teams wanting speaker labels or analytics from the same vendor
Not stated by the vendorPublished per minute for transcription models; code and weights carry the MIT licence
Pay-per-minute recognition with volume tiers and medical models
Ranked #10 of 12 in Best Speech-to-Text APIs in 2026.
Free tierPublished pricingElsewhere
The model list includes default, command_and_search, latest_short, latest_long, phone_call, video, medical_conversation and medical_dictation, with chirp in v2 only. Dynamic batch recognition in v2 costs $0.003 per minute, with every request rounded up to a whole second.
Savings plans of one or three years and Contact sales discounts exist for committed buyers. The cheapest standard rate needs two million minutes a month, so small accounts pay near the top of the scale.
What stands out
Volume price tiers
Medical models
Chirp model (v2)
Where it costs you
Medical models cost several times the standard minute rate.
v1 and v2 carry different rates, so the generation you call changes the bill.
Right for
High-volume buyers whose monthly minutes will reach the cheapest volume tiers
Wrong for
Small teams wanting one flat rate with no tiers
Not stated by the vendorPublished per minute with volume tiers; free tier covering the first 60 minutes monthly; sales quotes for discounts
Per-second billed transcription with call analytics and redaction add-ons
Ranked #11 of 12 in Best Speech-to-Text APIs in 2026.
Free tierPublished pricingElsewhere
Add-on rates in US East run per minute: PII redaction $0.0024, generative call summarisation $0.0024 and Call Analytics $0.0300 for the first 250,000 minutes, with Transcribe Medical near $0.075. Custom language models and toxicity detection are add-ons too.
Because rates change by region and volume discounts go through AWS sales, the published page gives only examples. Post-call and real-time analytics are separate capabilities from plain transcription.
What stands out
Per-second billing
Call Analytics
PII redaction
Where it costs you
The free tier lasts 12 months and covers 60 minutes a month, with no roll-over.
Redaction and custom language models fall outside the free allowance.
Right for
Buyers already on AWS who want call analytics beside transcription
Wrong for
Teams comparing prices without a region in mind
Not stated by the vendorPublished per second of audio in one-second steps; free tier for 12 months; add-ons per minute
Hourly speech billing with commitment tiers and disconnected containers
Ranked #12 of 12 in Best Speech-to-Text APIs in 2026.
Free tierPricing on requestNorth America
Commitment plans of 2,000, 10,000 and 50,000 hours a month apply to Standard and Custom, with an overage rate per hour, and the same three tiers cover the add-ons and connected containers. Disconnected container allowances reach 120,000 or 600,000 hours a year.
Continuous language identification, diarization and pronunciation assessment are priced add-ons, while single language identification is included. One custom endpoint model per month is free on F0.
What stands out
Commitment tiers
Disconnected containers
Real-time and batch
Where it costs you
No hourly amount could be recorded because the price fields showed placeholders.
Batch transcription is not supported on the F0 tier.
Right for
Organisations on Microsoft agreements needing commitment tiers or disconnected containers
Wrong for
Buyers wanting to compare exact hourly amounts from the page
United StatesPay-as-you-go per audio hour with monthly commitment tiers; the rates did not render on the page read
A speech-to-text API receives audio files or live streams over a programming interface and returns transcripts with timestamps, speaker labels and optional analysis, billed by audio duration or credits. Feature lists look alike, so we would settle these points one at a time instead.
01
Decide whether you need a published price
2 vendors on this page sell by quote, while 10 publish a pricing model you can read today: Soniox, AssemblyAI, Speechmatics, Deepgram, Vosk, Rev AI, ElevenLabs Scribe, OpenAI Whisper, Google Speech-to-Text, Amazon Transcribe.
02
Look at where the vendor sits
Check the vendor's home market (Elsewhere 10, North America 1, Europe 1), because that is where subpoenas and data requests arrive first, whatever region holds your files.
03
Decide if open source matters
Source code is available for Vosk. Running it yourself swaps the licence fee for your own upkeep.
Reading a speech-to-text API price in hours, minutes and credits
The first trap in this market is the unit. Soniox and AssemblyAI quote per audio hour, OpenAI Whisper and Google Speech-to-Text per minute, and ElevenLabs Scribe hides transcription inside a credit pool metered at 330 credits per minute. Convert everything to the same hour before comparing.
Then add what the base rate leaves out: AssemblyAI prices speaker identification, translation and entity detection per hour on top, whereas Soniox bundles diarization and translation. Amazon Transcribe adds PII redaction and Call Analytics per minute, and Azure Speech add-ons such as continuous language identification are charged per hour. Finally check rounding, because Amazon bills in one-second steps and Google rounds each request up to the nearest second.
Convert every quote to cost per audio hour
List the add-ons your job needs and price each
Check the rounding rule per request
Streaming against recorded audio and what the concurrency caps allow
Live transcription costs more than recorded audio at nearly every vendor listed here, so the split of your traffic drives the bill. Deepgram charges Nova-3 streaming at about 80% above the pre-recorded minute rate, Speechmatics lists Melia 1 at $0.24 an hour recorded and $0.40 real-time, and Soniox moves from $0.10 to $0.12. Concurrency then decides whether a launch day works.
AssemblyAI caps new streams at 5 a minute on the starter allowance and 100 once paying, Speechmatics gives its Free tier 2 concurrent real-time sessions, and Deepgram allows 150 streaming connections on its entry plan. Ask each vendor what happens when you exceed the cap, and whether raising it needs an Enterprise contract or only a support ticket.
Split expected hours into live and recorded
Read the concurrency cap for your peak hour
Ask whether raising the cap needs a contract
Cloud, container or on-device: where speech audio is allowed to run
Voice recordings often hold names, diagnoses and card numbers, so the deployment route can end a shortlist before price does. Vosk runs offline on a Raspberry Pi, phone or server and costs nothing to licence, but you carry the operations. Speechmatics, Deepgram and Rev AI offer containers or on-prem deployment, though Speechmatics and Deepgram keep containers in the Enterprise tier.
Voxist lists a European cloud, a customer VPC, a customer data centre and fully air-gapped sites. Azure Speech sells disconnected containers with yearly hour allowances. Cloud-only APIs such as Soniox and AssemblyAI instead offer regional processing, so ask which region holds your audio and whether a model can be run without sending it anywhere.
Decide whether audio may leave your network
Confirm containers are included in your tier
Ask which region processes and stores audio
Testing speech-to-text accuracy on your own recordings before committing
Vendor accuracy figures are hard to compare, and Voxist's 4.2% French word error rate comes with no method. Run the same sample of your own calls through several services instead. Starting credits make this cheap: Deepgram gives $200, Speechmatics $100, AssemblyAI $50 and Rev AI 5 hours of Reverb credit.
Google Speech-to-Text covers 60 minutes a month and Amazon Transcribe the same for 12 months, while ElevenLabs Scribe's 10k credits stretch to roughly half an hour. Soniox mentions no trial, and OpenAI Whisper lists none for the API, though its MIT-licensed model can be tested locally. Include noisy audio, regional accents and your own product names, and score the errors by hand, word by word.
Run one shared sample through three services
Include noisy audio, accents and product names
Score the errors yourself, not the vendor
What goes wrong most often when buying speech-to-text apis software
Comparing a per-minute Whisper price with a per-hour Soniox price without converting both to the same unit.
Budgeting from the base transcription rate and forgetting the per-hour add-ons for diarization, redaction or analytics.
Assuming containers are included, when Speechmatics and Deepgram reserve them for Enterprise contracts.
Trusting a vendor's quoted error rate, like Voxist's French figure, instead of testing recordings with your own accents.
07
Frequently asked questions
8 answers
What is the best speech-to-text apis in 2026?
First of 12 on this page is Soniox. Soniox prices transcription per audio hour, with recorded audio cheaper than real-time streaming, and folds speaker detection, language identification and smart formatting into that rate.
Real-time calls include translation at no extra cost. Weak spots: the pricing page mentions no trial or starter credit, the pages describe only a cloud API, and no country or owner is stated.
Which speech-to-text apis tools publish their pricing?
The ones that do, with the model each states:
Soniox: Published per audio hour, with asynchronous and real-time rates; enterprise discounts on request.
AssemblyAI: Published per audio hour by model; starter credit at signup; enterprise volume deals by contract.
Speechmatics: Published per hour by model tier, with subscriptions, credit packs and Enterprise contracts.
Deepgram: Published per minute with no minimum; prepaid annual Growth credits; Enterprise by contract.
Vosk: Free, Apache-2.0 open source code; the vendor page lists no cloud service or prices.
Rev AI: Published per hour or minute by product; starter credit of 5 hours; Enterprise by conversation.
ElevenLabs Scribe: Monthly subscriptions with a free plan and a shared credit pool; Enterprise priced by contract.
OpenAI Whisper: Published per minute for transcription models; code and weights carry the MIT licence.
Google Speech-to-Text: Published per minute with volume tiers; free tier covering the first 60 minutes monthly; sales quotes for discounts.
Amazon Transcribe: Published per second of audio in one-second steps; free tier for 12 months; add-ons per minute.
2 vendors do not print a price.
Is there a free speech-to-text apis tool?
Vosk, ElevenLabs Scribe, Google Speech-to-Text, Amazon Transcribe, Azure Speech offer a free tier or a free self-hosted edition.
Where are these speech-to-text apis vendors established?
Vendor countries on this page, Elsewhere 10, North America 1, Europe 1:
Soniox: Not stated by the vendor.
AssemblyAI: Not stated by the vendor.
Speechmatics: Not stated by the vendor.
Deepgram: Not stated by the vendor.
Vosk: Not stated by the vendor.
Rev AI: Not stated by the vendor.
Voxist: France.
ElevenLabs Scribe: Not stated by the vendor.
OpenAI Whisper: Not stated by the vendor.
Google Speech-to-Text: Not stated by the vendor.
Amazon Transcribe: Not stated by the vendor.
Azure Speech: United States.
Which speech-to-text apis tools are open source?
Vosk.
Which speech-to-text apis tools can you host yourself?
Speechmatics, Deepgram, Rev AI. 9 products offer no self-hosted option.
What should you use instead of Soniox?
AssemblyAI and Speechmatics are the next two on this page. AssemblyAI is for Developers who want to start on credits and pick analysis features one by one; Speechmatics is for teams wanting model choice by price tier plus a route to on-device deployment.
Who should not buy Soniox?
Buyers who must run recognition inside their own data centre. Testing means paying from the first hour, since nothing on the page offers a sample allowance..
Missing from this speech-to-text apis ranking? Contact us with a short description of the product and its users. Adding it is an editorial choice, and software advice explains what a listing does and does not involve.
