European Alternatives to Deepgram
Looking for a European alternative to Deepgram? Deepgram sells speech-to-text, text-to-speech and voice agent APIs, and its terms of service are governed by California law, with disputes settled by individual arbitration or in the state and federal courts in San Francisco. Its pricing page lists a free $200 credit, then pay-as-you-go, and prepaid Growth plans from $4,000 a year that save up to 20%.
The European Speech-to-Text API tools below contract from the UK, France, Germany, the Netherlands, Denmark and Romania, publish per-minute or per-hour rates, and several state EU-only processing or on-premise deployment.
How we rank these tools — 4-step process
-
1
European ownership, verified
The company is headquartered and incorporated in the EU, EEA, the UK, Switzerland or Israel, and processes customer data in Europe or Israel. A US parent company disqualifies a tool from this page regardless of where its servers are.
-
2
Category fit and hands-on review
What the tool actually does, who it suits, and where it falls short — checked against the vendor’s own documentation, changelog and pricing page rather than its marketing copy.
-
3
Compliance and pricing check
GDPR posture, hosting location and the prices quoted on this page are verified against the vendor’s public pricing before publication, and re-checked when we revisit the category.
-
4
Position on this page
Placement on this page can be paid, and that can affect which tools appear here and the order they appear in. It never buys a good review: a tool that fails the checks above is not here at any price, and payment does not change the shortcomings we write about. A vendor can ask us to correct a factual error — not to remove a criticism.
Vendors can pay for visibility on this page. It never changes what an entry says about a product, including the criticism, and we earn nothing when you click through to a vendor. Vendors can pay for more visibility, higher up on the review pages; it never affects which tools are listed. Editorial policy
12 European Alternatives to Deepgram
Speechmatics
Cambridge speech API with published hourly rates and free credits
Gladia
French real-time and async speech API without subscriptions
Voxtral
Mistral's open-weight speech models, billed per minute on the API
Reson8
Amsterdam speech API that adapts to your vocabulary from text
Voxist
Paris speech API priced in euros with a free tier
DeepScript
Pay-per-use transcription hosted on Hetzner servers in Germany
Vatis Tech
Romanian speech API with async rates per hour
Corti
Copenhagen clinical speech API with EU or US hosting choice
Phonexia
Brno on-premise speech recognition and voice biometrics platform
Scriptix
Hague speech platform with subscriptions and air-gapped on-premise
Vocapia
Orsay research company with the VoxSigma speech software suite
pyannoteAI
French diarization API that adds speaker labels to transcripts
Key takeaways
- Deepgram's terms are governed by California law, with disputes settled by individual arbitration or in San Francisco courts.
- Deepgram's privacy policy says data is stored on servers in the U.S. and that Standard Contractual Clauses cover transfers out of the EEA.
- Twelve European speech-to-text vendors on this page contract from the UK, France, Germany, the Netherlands, Denmark, Romania and the Czech Republic.
- Voxtral publishes the lowest per-minute rate on the page at $0.0035, and Reson8 advertises a banner price from €0.01 per hour for prerecorded audio.
- Speechmatics, Phonexia, Scriptix, Vatis Tech and Vocapia can run on your own infrastructure, which Deepgram's pages we read do not describe.
Why people leave Deepgram
Deepgram is a California company that sells speech-to-text, text-to-speech and voice agent APIs, plus an audio intelligence layer on top of the transcripts. Its home page lists contact centres, restaurants, pharma, medical and government as target industries, and a banner announces that Daily and Pipecat are joining Deepgram. A developer who starts with transcription on Deepgram can later add synthesis and an agent runtime without opening another vendor account.
The legal side is where European buyers hesitate. Deepgram's terms of service are governed by the law of California, individual arbitration replaces class actions, and claims that cannot be arbitrated go to the state or federal courts in San Francisco. Its privacy policy says Deepgram stores data on servers in the U.S. and relies on the EU Standard Contractual Clauses with customers to move Customer Data out of the EEA, the UK and Switzerland.
None of this makes Deepgram unusable in Europe, and many companies sign exactly these terms. It does mean that a European customer buys a transcription service whose default data path ends on American servers under American courts. A data protection officer who has to explain that path in a processing register, or a public body that cannot send call recordings across the Atlantic, has a concrete reason to look at speech-to-text vendors in Cambridge, Paris, Amsterdam or Brno.
- The default data path ends on servers in the United States Deepgram's privacy policy states that it stores data on servers in the U.S., and that it uses the Standard Contractual Clauses to legitimise transfers of Customer Data from the EEA, the UK and Switzerland. That is a lawful mechanism, but it is a transfer. Every recording you send is audio of real people, often with names, account numbers or health details spoken aloud, so the transfer is the part a review board asks about first.
- Disputes go to California, and credits are not refunded The terms apply California law and send most disputes to individual arbitration, with other claims heard in San Francisco. They also state that unused time or credits on a subscription or pre-payment are not refunded. A company in Utrecht or Lyon that pre-pays the Growth plan, which starts at $4,000 a year, is therefore committing both cash and a foreign venue.
- Prepaid commitments pull you towards one vendor Deepgram pitches pay-as-you-go with a $200 free credit, then offers Growth with pre-paid credits and savings of up to 20%. The saving is real, yet it only exists if you stay on Deepgram for the year. A team that suspects its recordings will need to stay in Europe within twelve months has little reason to lock the discount in first.
- The platform keeps widening beyond transcription Deepgram now spans transcription, synthesis, voice agents and audio intelligence, and its home page announces that Daily and Pipecat are joining. If you only wanted a transcript endpoint, you are buying into a roadmap that points elsewhere. Several European vendors stay narrow on purpose, which makes their scope, terms and invoices easier to read.
What you have to replace, not just match
The first thing to replace is the transcription endpoint itself, and that part is small. Most European speech APIs here accept audio over REST, and Voxist adds gRPC with SDKs for Python, Node.js, Go, Rust, Java and .NET. Swapping the endpoint means changing the client, the authentication and the response mapping for words, timestamps and speaker labels, which is usually a few days of work plus a re-run of your test recordings.
The second thing is everything that sat behind the endpoint. If your product uses Deepgram for live streaming into a voice agent, you must check streaming latency, partial results and end-of-speech detection on the new vendor, because those behaviours differ per provider. If you also use Deepgram text-to-speech or its agent runtime, the European speech-to-text vendors below do not replace them, and only Voxist and Voxtral list synthesis among their services.
The third is your own vocabulary and evaluation set. Any custom terms, keyword boosts or domain tuning you built on Deepgram must be rebuilt, for example as the text list Reson8 uses or the custom vocabulary DeepScript, Voxist and Voxtral document. Collect a few hours of your real audio before you move, because vendor accuracy figures, including those in this article, say little about your accents and microphones.
The alternatives compared
| Position | Tool | Headquarters | Pricing | Jurisdiction |
|---|---|---|---|---|
| #1 | Speechmatics | Cambridge, United Kingdom | Free tier with $100 credit, no card required / Pro pay-as-you-go from $0.129 / Enterprise custom pricing with volume discounts | United Kingdom (adequacy decision, outside the EEA) |
| #2 | Gladia | Roubaix, France | Starter pay-as-you-go from $0.61/hr async, $0.75/hr real-time (€50 free credits) / Growth from $0.20/hr with commitment / Enterprise custom | EU (France) |
| #3 | Voxtral | Paris, France | API: Voxtral Mini Transcribe 3 $0.0035 per minute, Voxtral Small Transcribe 3 $0.004 per minute, Voxtral Mini Transcribe Realtime 3 $0.008 per minute (Mistral pricing page, USD or EUR toggle) | EU (France) |
| #4 | Reson8 | Amsterdam, Netherlands | Prerecorded transcription in 40+ languages 'from EUR 0.01 per hour' (home page banner for Rumble-2-Flash); full rate card not read | EU (Netherlands) |
| #5 | Voxist | Paris, France | Free: EUR 0 a month with 5 hours of audio; Pay-as-you-go: EUR 0.006 per minute billed per second; Enterprise: custom | EU (France) |
| #6 | DeepScript | Mannheim, Germany | Pay per use: Standard $0.20 per hour, Premium (Topaz) $0.30 per hour, Premium (Jade) $0.32 per hour; Pro $22.00 per month; 3 free transcriptions | EU (Germany) |
| #7 | Vatis Tech | Vădăstrița, Olt, Romania | Free 10 minutes; Basic €10/month (10 h), Standard €27/month (30 h), Pro €89/month (100 h); extra hours €5; API from €0.22/hour | EU (Romania) |
| #8 | Corti | Copenhagen, Denmark | Speech to Text $0.0065 per minute; free credits across capabilities; enterprise plan on request | EU (Denmark) |
| #9 | Phonexia | Brno, Czech Republic | Pricing on request | EU (Czech Republic) |
| #10 | Scriptix | The Hague, Netherlands | Bronze EUR 119 a month, Silver EUR 479 a month, Gold EUR 1,199 a month (annual plans paid upfront); on-premise from EUR 1,999 per instance | EU (Netherlands) |
| #11 | Vocapia | Orsay, France | On the order of €0.01 per minute for generic systems in large quantities (vendor statement); no pay-as-you-go plan | EU (France) |
| #12 | pyannoteAI | Neuilly-sur-Seine, France | Developer EUR 19 a month with EUR 19 usage credit; Starter EUR 99 a month with EUR 99 credit; Enterprise custom; 30-day trial | EU (France) |
How each alternative compares to Deepgram
Best for: Teams that want live and batch transcription from a UK vendor and may need offline deployment
Speechmatics, trading as Cantab Research Limited in Cambridge, sells a pure speech-to-text API. Against Deepgram it covers 55+ languages with speaker diarization and sub-second real-time results, and it publishes ISO 27001, GDPR, HIPAA and SOC 2 Type II as certifications. A Medical Model targets clinical vocabulary, which parallels Deepgram's own medical focus.
The contractual picture differs from Deepgram's in one respect: Speechmatics can run in a multi-region cloud, on-premise or on-device. The company is British, so the relationship with the EU rests on an adequacy decision rather than establishment in a member state. Pro pay-as-you-go starts at $0.129, with automatic volume discounts above 500 hours a month and a $100 free credit.
What Speechmatics does better than Deepgram
- Offers on-premise and on-device deployment, which Deepgram's pages we read do not describe
- Lists its certifications on its own pages, including SOC 2 Type II and ISO 27001
- Reports an independent benchmark in which it had the lowest error rate of twelve services
- Applies volume discounts automatically above 500 hours a month, without a pre-paid year
Where Speechmatics is a step down from Deepgram
- No text-to-speech, voice cloning or agent runtime, so it covers only part of Deepgram's range
- Smaller free credit ($100 against Deepgram's $200)
- UK establishment means adequacy rather than EU membership
Standout against Deepgram. Speechmatics is the only API here that combines a published accuracy benchmark with cloud, on-premise and on-device deployment.
Best for: Product teams that want a hosted EU speech API billed per hour instead of per contract
Gladia SAS in Roubaix is part of OVHcloud and sells transcription as an API with asynchronous and real-time modes. Language detection runs automatically across 100+ languages, and diarization is included in the standard offer. Its Enterprise tier adds zero data retention, dedicated infrastructure and an SLA, which answers the questions European compliance teams usually put to a US vendor.
Billing is usage-based: Starter is $0.61 per hour for async and $0.75 for real-time, with €50 of credit that does not expire, and Growth starts at $0.20 per hour on commitment. Gladia is cloud only, so it cannot help if recordings must stay inside your own network. The company also publishes terms for a Delaware entity, Gladia Inc., next to those of Gladia SAS.
What Gladia does better than Deepgram
- Sits inside a European cloud group, OVHcloud, with the French register showing an OVH entity as president
- Enterprise tier states zero data retention and dedicated infrastructure
- Starter plan needs no pre-payment, and the €50 credit has no expiry
- Lists SOC 2 Type II, GDPR and HIPAA on its own pages
Where Gladia is a step down from Deepgram
- Cloud only, with no on-premise or on-device edition
- No text-to-speech or voice agent product
- A Delaware affiliate publishes terms too, so read which entity you contract with
Standout against Deepgram. Gladia is the one hosted API here owned by a European cloud group, which gives its infrastructure story a named parent.
Best for: Developers who want a per-minute API now and the choice to host the weights later
Voxtral is Mistral AI's speech model line, built by a Paris company registered under RCS 952 418 325. The API lists Mini Transcribe 3 at $0.0035 per minute, Small Transcribe 3 at $0.004 and a Realtime model at $0.008. Documented features include live transcription, speaker labels, timestamps and custom vocabulary, and the same API also covers text-to-speech.
Against Deepgram, the unusual part is the licence: the original 24B and 3B models are released under Apache 2.0, so a company can run them on its own GPUs. The cost is clarity about the hosted service. The pages we read do not say where API audio is processed, and Mistral's own website is hosted by Netlify in San Francisco, so confirm data location in your contract.
What Voxtral does better than Deepgram
- Publishes the lowest per-minute rate among the vendors here that quote prices per minute
- Original models are available as Apache 2.0 weights, an exit that Deepgram does not give
- French company with a published register number
- Transcription and text-to-speech share one API
Where Voxtral is a step down from Deepgram
- Hosted-API data location is not stated on the pages we read
- Benchmark claims come from Mistral itself
- No contact-centre analytics comparable to Deepgram's audio intelligence layer
Standout against Deepgram. Voxtral is the only entry here where the model weights are open, so leaving the vendor does not mean leaving the model.
Best for: Dutch and European audio full of drug names, case numbers or product codes
Reson8 is operated by Resonate Labs B.V. in Amsterdam, with Chamber of Commerce number 98891340. The company states that all GPU processing takes place in the EU and that audio is not retained. Its distinguishing feature is vocabulary adaptation from a plain text list, with no fine-tuning run, aimed at domain terms in Dutch, German, French and other European languages including Frisian.
The pricing line on the home page reads from €0.01 per hour for prerecorded transcription in 40+ languages, but the rate card itself was not readable as text, so streaming prices and free tiers are unconfirmed. The company is young, with terms and privacy policy updated in October 2026. No certifications or named hosting provider appeared on the pages we read.
What Reson8 does better than Deepgram
- States EU-only GPU processing and zero audio retention, where Deepgram says it stores data in the U.S.
- Adapts to custom terms from a text upload, without a training project
- Prices in euros, with a banner rate from €0.01 per hour
Where Reson8 is a step down from Deepgram
- No streaming or voice agent product described on the pages we read
- No certifications or named sub-processors visible, which procurement will ask about
- Full price list could not be verified
Standout against Deepgram. Reson8 turns a pasted word list into recognition vocabulary, which suits transcripts full of names and codes.
Best for: French and EU developers who want transcription, translation and synthesis behind one euro-billed API
Voxist SAS in Paris offers streaming and batch recognition in 40+ languages, translation, neural text-to-speech and a voice-to-voice endpoint. Access is by REST or gRPC, with SDKs for six languages. Voxist states that it hosts in EU data centres run by OVHcloud and Scaleway, so no US cloud sits in the stated chain.
Pricing is in euros: a free plan with 5 hours a month, then €0.006 per minute billed per second. An Enterprise plan adds on-premise and air-gapped options. SOC 2 Type II and ISO 27001 are listed as in progress, not complete, and the free plan is limited to 5 requests per second. The homepage names five recognition languages, so test Dutch or Nordic audio first.
What Voxist does better than Deepgram
- Hosts on OVHcloud and Scaleway in the EU, against Deepgram's stated U.S. servers
- Bills in euros and per second, with no pre-paid year
- Spans recognition, translation and synthesis, so it mirrors Deepgram's breadth better than most
- Offers an on-premise, air-gapped Enterprise option
Where Voxist is a step down from Deepgram
- Security certifications are still in progress
- Accuracy figures are the vendor's own and the homepage lists fewer languages than some rivals
- Free plan is rate limited and capped at 5 hours
Standout against Deepgram. Voxist is the only European vendor on this page that sells recognition, translation and speech output in euros from EU data centres.
Best for: German-speaking users who transcribe recordings and want an API, hosted at Hetzner in Germany
DeepScript is a brand of Aliru GmbH in Mannheim, entered at Amtsgericht Mannheim as HRB 739079. Transcription runs on Hetzner servers in Nuremberg. The service handles 99+ languages with DACH dialect recognition, diarization and custom vocabulary, exports TXT, SRT, VTT and JSON, and exposes a REST API, webhooks and an MCP endpoint for connecting AI assistants.
Rates are per use: Standard at $0.20 per hour, Premium Topaz at $0.30 and Premium Jade at $0.32, with a Pro plan at $22.00 a month and three free transcriptions. The feature list shows no real-time streaming, so it does not replace Deepgram in live use. Transcripts are deleted after 30 days unless you hold Pro, and the English page shows dollars with a euro toggle.
What DeepScript does better than Deepgram
- Names its hosting provider and city, where Deepgram's privacy policy only says U.S. servers
- Supports DACH dialects, which a general API rarely tunes for
- Prices by the hour with no minimum and no pre-paid contract
Where DeepScript is a step down from Deepgram
- No streaming in the feature list we read, so no voice agent use
- No text-to-speech, agent runtime or analytics suite
- Transcripts expire after 30 days without the Pro plan
Standout against Deepgram. DeepScript is the one entry here that names a German city and a German data centre operator for its servers.
Best for: Small companies wanting plan prices in euros and an API rate per hour, with possible self-hosting
Vatis Tech S.R.L. is registered with the Romanian Trade Registry under J28/696/2021. It runs an editor-based app in 50+ languages next to an asynchronous API at €0.22 per hour, including diarization and word timestamps, with audio intelligence billed per token on top. Plans are Basic €10 for 10 hours, Standard €27 for 30 and Pro €89 for 100.
Against Deepgram, the useful differences are published euro prices and the stated on-premise or isolated-cloud deployment. The weaker points are practical: hosted data location is not stated, SOC 2 Type II is unfinished (ISO 27001 is stated), accuracy comparisons are the vendor's own, and real-time Romanian is announced rather than shipped.
What Vatis Tech does better than Deepgram
- Shows every plan and the API rate in euros on one pricing page
- Offers on-premise or isolated-cloud deployment for regulated audio
- Gives you an editor and an API in one product
Where Vatis Tech is a step down from Deepgram
- Does not state where hosted customer data lives
- API is asynchronous, with live Romanian still upcoming
- A smaller vendor without Deepgram's voice agent or synthesis products
Standout against Deepgram. Vatis Tech shows its monthly plans and its API hourly rate together in one euro price list.
Best for: Health software vendors building clinical documentation on EU or US hosting that they choose
Corti ApS in Copenhagen (VAT number 35669825) offers real-time and batch clinical transcription, a medical coding API, text generation and an agentic framework. Speech to Text costs $0.0065 per minute, and the customer picks EU or US hosting. Deepgram also targets medical customers and its terms provide for a BAA, so the comparison is real for a health vendor.
The limits are scope and structure. The speech product is clinical, so general audio and non-medical languages are not documented. The privacy policy names US affiliates, Corti Inc. and Corti America Inc., and the ownership structure is not stated. Prices appear in dollars, and the API master terms were not read for this comparison.
What Corti does better than Deepgram
- Lets the customer choose EU hosting explicitly
- Bundles medical coding and note generation with transcription
- Danish company registered in Copenhagen
Where Corti is a step down from Deepgram
- Clinical scope, with no general-purpose or call-centre positioning
- Names US affiliates, which weakens a pure-EU story
- Dollar pricing and unread master terms
Standout against Deepgram. Corti pairs speech recognition with medical coding in the same platform, a combination none of the other vendors here describes.
Best for: Security agencies, law enforcement and call centres that keep audio inside their own network
Phonexia a.s. in Brno builds a modular platform of 14 speech technologies behind one API: transcription, speaker identification, voice biometrics and deepfake detection. The site states it is designed for on-premise and private-cloud deployment, and it names national security, law enforcement and call centres among its users. It advertises 87+ languages, a vendor figure.
That makes Phonexia a different kind of replacement for Deepgram: nothing is hosted by Phonexia unless you arrange it, so no audio crosses the Atlantic by default. The price is procurement friction. No prices are published, there is a demo rather than self-service sign-up, and no transcript editor or subtitle tooling is advertised.
What Phonexia does better than Deepgram
- Runs inside your network, so audio never needs to reach a vendor's servers
- Adds voice biometrics and deepfake detection that Deepgram's pages we read do not list
- Czech company whose privacy policy is governed by Czech law
Where Phonexia is a step down from Deepgram
- No published pricing and no free credit
- Needs a sales conversation and infrastructure of your own
- Language count is the vendor's own figure
Standout against Deepgram. Phonexia puts speaker biometrics and deepfake detection next to transcription in one on-premise platform.
Best for: Courts, public bodies and broadcasters that want editors, subtitles and an air-gapped option
Scriptix B.V. in The Hague (Chamber of Commerce 75668475) runs batch and realtime recognition inside a platform with transcript and subtitle editors, machine translation in 100+ languages, meeting minutes and custom model tuning. Data stays on EU infrastructure by default, or inside your own tenant on-premise. The team is four people.
Pricing is subscription based: Bronze €119 a month, Silver €479 and Gold €1,199 on annual plans paid upfront, and on-premise from €1,999 per instance. That model suits fixed workloads and fits poorly for spiky API traffic billed per minute, as with Deepgram's pay-as-you-go. Allowances are described as fair use rather than in hours, and the language count is 50.
What Scriptix does better than Deepgram
- Starting prices for on-premise and air-gapped editions are public
- Includes editors, subtitles and translation in the same product as the recogniser
- Defaults to EU infrastructure
Where Scriptix is a step down from Deepgram
- Subscription tiers, not metered per minute, so costs are flat
- Fewer languages (50) than several API-first rivals
- Very small team, which matters for support continuity
Standout against Deepgram. Scriptix is the only vendor here that publishes a starting price for an air-gapped on-premise edition.
Best for: Archive owners and broadcasters transcribing many hours in bulk
Vocapia Research SAS is registered in Evry and based at Parc Orsay Université. Its VoxSigma suite does speech-to-text, language identification, speech-text alignment, translation and summaries, delivered as standalone software or as a REST web service that returns XML or JSON with diarization and confidence scores. The service lists 28 languages, including Dutch, French, German and Polish.
The pricing model is the sharpest contrast with Deepgram. Vocapia states about €0.01 per minute for generic systems in large quantities, billed on speech duration, but it no longer offers pay-as-you-go and sends small users to a partner service. Hosting regions are not described, and the tooling is developer oriented.
What Vocapia does better than Deepgram
- Sells standalone software, so processing can stay on your servers
- States a low volume price of about €0.01 per minute
- French company registered in Evry
Where Vocapia is a step down from Deepgram
- No pay-as-you-go plan and no free credit
- 28 languages, a shorter list than larger API providers
- Price is an order-of-magnitude statement, not a rate card
Standout against Deepgram. Vocapia is the only vendor here that states a volume price per minute while refusing self-service small accounts.
Best for: Teams with an existing transcription engine that need accurate speaker turns in calls and meetings
pyannote SAS (pyannoteAI) in Neuilly-sur-Seine, registered at Nanterre under 924 837 636, focuses on speaker diarization: who spoke when. It offers streaming and batch diarization, voice activity detection, speaker identification and separation, and speaker-attributed transcripts. It also states you can use the open source library, the API or your own infrastructure.
This only replaces the diarization slice of Deepgram. The speech engine behind its transcripts is not named on the pages we read, nor is data location or the open source licence. Plans are Developer at €19 a month with €19 credit, Starter at €99 with €99 credit, and Enterprise on request, with a 30-day trial.
What pyannoteAI does better than Deepgram
- Specialises in speaker turns, which a general API treats as an add-on
- Lets you run the open source or your own infrastructure option
- Publishes euro plans from €19 a month
Where pyannoteAI is a step down from Deepgram
- Not a full speech recogniser, so it cannot stand alone
- Data location and speech engine are not stated
- Starter plan includes only three users
Standout against Deepgram. pyannoteAI is built around one task, speaker diarization, and offers it as open source, an API or on your own hardware.
What actually breaks when you switch
The hardest part of leaving Deepgram is rarely the code. A transcription API behaves differently on your own audio: punctuation style, number formatting, how overlapping speakers are split, and how long a stream waits before closing a sentence. Run the same two hundred recordings through the old and new vendor, diff the output, and decide which differences your downstream parsers, summaries or agents will actually notice.
Contracts are the second snag. If you pre-paid a Growth plan, the terms say unused credits are not refunded, so schedule the switch around the renewal date and use the remaining credit on non-critical jobs. Keep a fallback during rollout, for example by routing a share of traffic to the new vendor first. Also update the processing register, sub-processor list and privacy notice, since the recipient of your audio changes.
Is Deepgram allowed under GDPR for European call recordings?
Deepgram can be used under GDPR if the contract and transfer mechanism are in place, and the vendor itself says it signs Standard Contractual Clauses with customers for exactly this purpose. The decision is therefore less about legality in the abstract and more about your risk appetite. You remain the controller, so your processing register must record that recordings reach servers in the U.S.
For ordinary customer service audio, many companies accept that. For recordings of patients, citizens or employees, a European vendor with EU processing removes a whole section of the transfer impact assessment. Voxist (OVHcloud and Scaleway), DeepScript (Hetzner in Germany) and Reson8 (EU GPUs, zero audio retention) each state EU processing, and Gladia, Scriptix and Corti offer EU options.
Which European speech API is closest to Deepgram for real-time voice agents?
Speechmatics and Gladia are the closest on streaming speech recognition, since both offer real-time transcription alongside batch. Voxtral has a dedicated realtime model at $0.008 per minute, and Voxist documents streaming over gRPC. Corti streams too, but only for clinical conversations.
None of these replaces Deepgram's full voice agent runtime, so a team building an agent will still assemble synthesis and orchestration separately. Test end-of-speech detection and first-word latency with your own audio before you commit. DeepScript and Reson8 are not candidates for live work: DeepScript lists no streaming and Reson8's banner price is for prerecorded files.
Which one to pick
For a like-for-like hosted API, start with Speechmatics, Gladia and Voxtral, test them on your own audio, and pick on latency, language coverage and the price you can actually read. Choose Voxist when billing in euros and EU data centres matter more than breadth of languages, and Reson8 when your audio is full of domain terms. If you need streaming plus text-to-speech, only Voxist and Voxtral list synthesis.
For regulated or sensitive audio, choose the vendors that can sit inside your network: Phonexia, Scriptix, Vocapia, Vatis Tech or Speechmatics on-premise. Pick Corti for clinical documentation and DeepScript for German recordings on Hetzner hardware. Add pyannoteAI only if speaker labels are your weak spot. Stay on Deepgram if you depend on its voice agent runtime and your legal team has already accepted the U.S. data path.
Frequently Asked Questions
For a general API with published rates, Speechmatics (Cambridge), Gladia (Roubaix) and Voxtral (Paris) are the strongest starting points, because they cover streaming and batch and list prices openly. Voxist and Reson8 suit euro billing and EU processing. Phonexia and Scriptix suit on-premise needs. Which is best depends on your criterion: price, deployment, or language coverage.
On the rates we could read, Voxtral charges $0.0035 per minute for Mini Transcribe, Voxist €0.006 per minute and Corti $0.0065 per minute. Reson8 advertises prerecorded audio from €0.01 per hour. We could not read Deepgram's rate card as text, so compare these figures with your own Deepgram invoice rather than with our summary.
Yes, with several vendors. Speechmatics offers on-premise and on-device deployment, Phonexia is built for on-premise and private cloud, Scriptix sells air-gapped on-premise from €1,999 per instance, Vatis Tech offers on-premise or an isolated cloud, and Vocapia sells VoxSigma as standalone software. Voxtral's original models are released as Apache 2.0 weights you can host yourself.
Deepgram gives a $200 credit, then pay-as-you-go with no minimums or expiry, according to its pricing page. European offers are smaller: Speechmatics gives $100 in credit, Voxist 5 hours a month, Gladia €50 in credits, pyannoteAI a 30-day trial and Vatis Tech 10 minutes. DeepScript allows three free transcriptions.
We did not verify Deepgram's language list, so no direct comparison is possible here. Among European vendors the stated counts range from 28 languages at Vocapia and 50 at Scriptix to 99 or more at DeepScript and more than 100 at Gladia. Counts are vendor figures, and accuracy for a given language still needs a test with your audio.
Related comparisons
About the author
Explore More European Alternatives
Discover privacy-focused European alternatives to other popular US tech services.