Gladia
Gladia is a single speech-to-text API for developers and product teams: asynchronous and sub-300 ms real-time transcription in 100 languages, with diarization, translation, summarization and entity extraction included, plus European data residency from a French publisher.
What is Gladia?
Gladia is an end-to-end audio infrastructure layer: record, transcribe and enrich audio through a single API. It is published by GLADIA SAS, founded in 2022 and based in Rennes, France, with a second entity, Gladia Inc., in Delaware for US contracts.
The product splits into two API families and one enrichment layer. Async STT handles recorded files and long-form content over REST; Real-time STT streams over WebSocket with sub-300 ms latency; and the Audio Intelligence layer adds diarization, translation, summarization, chapters, sentiment, named-entity recognition and PII redaction, included rather than billed as separate add-ons. Behind them sit Gladia's own Solaria models: Solaria-1 for universal multilingual coverage and Solaria-3, credited with 9.6% WER on real-world English audio and gains on English, French, German, Spanish and Italian.
Coverage is the headline differentiator: 100 languages, accent-aware automatic language detection and native code-switching, so speakers can change language mid-sentence without breaking the transcript. The home page claims 2 billion+ minutes transcribed, 300,000+ developers, 2,000+ enterprise teams and a 99.95% uptime SLA.
The second differentiator is sovereignty. Audio is processed by default on a European provider located in France, EU data residency is advertised at 100%, and EU and US workloads run on separate regional infrastructure. Compliance claims cover GDPR, HIPAA, SOC 2 Type II and ISO 27001, with a Trust Center and a public bug bounty.
For builders, Gladia ships Python and Node.js SDKs, WebSocket streaming, a CLI and 13 documented integrations, among them Pipecat, LiveKit, Twilio, Vapi, Recall.ai, Meeting BaaS, Attendee, VideoSDK, Composio, Zapier, Make and n8n. Four use cases are pushed forward: meeting assistants, CCaaS, voice agents, and content and media. Named customers include Aircall, VEED, Livestorm, Recall, Daily, Carv, Attention, Mojo and ScorePlay.
Gladia also publishes open benchmarks against nine providers over eight datasets and 74+ hours of audio, with an open-source methodology on GitHub, and maintains GladiaFlow, an MIT-licensed desktop dictation app for macOS and Windows that consumes the same API. The scope is deliberately one-directional: speech to text, never text to speech.
What it does
- Transcribe recorded audio and video files asynchronously through a REST API
- Transcribe live audio in real time over WebSocket with sub-300 ms latency
- Identify who speaks and when, with speaker diarization and word-level timestamps
- Detect the language and switch automatically mid-conversation across 100 languages
- Translate the transcript and export SRT or VTT subtitles from the same request
- Extract named entities, sentiment, chapters and summaries, and redact personal data
- Run custom LLM prompts on the transcript and push results to a CRM, database or warehouse via webhooks, Zapier, Make or n8n
When to use Gladia / When not to
A quick filter to help you decide if Gladia is the right fit.
When to use Gladia
- Engineering teams building meeting assistants, voice agents, CCaaS platforms or media tooling that want one API for capture, transcription and enrichment
- Product teams shipping multilingual apps where speakers switch languages mid-sentence, thanks to 100-language coverage and native code-switching
- European buyers under sovereignty or compliance constraints (GDPR, HIPAA, SOC 2 Type II, ISO 27001) who need EU data residency and a published DPA
- Teams moving off a self-hosted Whisper setup to cut scaling and DevOps effort, with a dedicated TCO calculator and documented migration guides
- High-volume workloads that need diarization, sentiment and named entities included in the rate rather than chained across several vendors, with volume discounts on the Growth and Enterprise tiers
When not to use Gladia
- Anyone needing text-to-speech or voice cloning: Gladia only converts audio into text, never the reverse
- Non-technical end users looking for a ready-made application, since the core product runs through an API key and code
- Mobile-first users: there is no iOS or Android app, and the only installable software is GladiaFlow, a desktop dictation tool for macOS and Windows
- Teams that require a permanently free tier, because the €50 credit is a one-time grant with no monthly reset
- Buyers who need public pricing at every level, as only the Starter tier is priced publicly while Growth and Enterprise go through sales
How to use Gladia
A typical end-to-end flow, from setup to results.
- Create an account on app.gladia.io/signup; signing up unlocks €50 in transcription credits
- Collect your API key from the Developer Console
- Try it without code in the Playground by uploading a file or starting a live session
- For recordings, POST the file to /v2/upload, then /v2/pre-recorded/init, and fetch the result by its id or receive it on a webhook
- For live audio, open a session with /v2/live/init, then stream chunks over WebSocket and consume the events
- Enable the add-ons you need in the request parameters: diarization, translation, sentiment, summary, named-entity recognition, PII redaction
- Use the official Python or Node.js SDK, or the Gladia CLI to transcribe from the terminal
- Plug into a prebuilt integration for voice-agent frameworks (Pipecat, LiveKit, Vapi, Retell) or meeting bots (Recall.ai, Meeting BaaS, Attendee)
- Automate without code through Zapier, Make or n8n, and follow the documented migration guides if you are coming from Deepgram or AssemblyAI
- Fund the credit wallet manually or enable auto top-up below a threshold of your choice; Gladia claims under 4 hours on average from first call to production traffic
Pros & Cons
Pros
- Broad language coverage, 100 languages with native code-switching, where the home page comparison table puts several rivals far lower (Deepgram 30+, AssemblyAI 6 languages in real time)
- Enrichment included in the price: diarization and named entities are listed as "Included" for Gladia and as "Add-on ($)" or "Limited" for several competitors
- Training opt-out applied by default, where three competitors in the same table charge for it
- EU data residency, with audio processed by default on a European provider located in France
- Complete, documented compliance stack: GDPR, HIPAA, SOC 2 Type II, ISO 27001, a publicly available DPA and BAA, and Annex III of the DPA naming all seven subprocessors with their data locations
- Open benchmarks against nine providers with an open-source methodology on GitHub, regressions included
- Fast onboarding: integration claimed in under 4 hours, migration guides from Deepgram and AssemblyAI, usage-based pricing with no setup fee or hidden cost per the FAQ, a non-expiring €50 trial credit worth roughly 80 h async or 60 h real time, and a 4.8 G2 rating displayed on the pricing page
Cons
- No permanently free plan: the €50 credit is a one-time grant with no monthly reset
- Only the Starter tier has a public price; Growth and Enterprise require a sales conversation
- The /security page has not been updated since 09/02/2024 and still describes "Free" and "Pro" plans that no longer exist, with a model-training rule that contradicts the Compliance Hub
- The same page presents ISO 27001 as work in progress while the home page and footer announce it as obtained
- Compliance documents (SOC 2 and HIPAA reports, penetration test, up-to-date subprocessor list) are only available on request through the Trust Center
- No mobile app, GladiaFlow is limited to macOS and Windows and is downloaded from GitHub rather than an app store, and there is no text-to-speech: the tool covers only one direction of the voice chain
- Custom hosting and on-premises deployment are reserved for the Enterprise tier, the home page comparison table is produced by Gladia against its own criteria, and the og:default images referenced by several pages return a 404
Pricing & Plans
There is no permanently free plan. New accounts receive €50 in transcription credits, a one-time grant with no expiry and no monthly reset, worth roughly 80 hours of pre-recorded transcription or 60 hours of real-time transcription. Beyond that, Gladia is billed on usage through a prepaid credit wallet, topped up manually or automatically. The lowest published entry point is the Starter tier at USD 0.61 per hour of asynchronous transcription, and USD 0.75 per hour in real time. Growth goes down to USD 0.20 per hour asynchronous and USD 0.25 per hour real time, up to 67% below Starter, against an upfront commitment; Enterprise is quoted on request, on an annual plan with unbundled pricing and volume discounts. Payment is handled through Stripe with Visa and Mastercard, while bank transfer and invoicing are available at enterprise level. The FAQ states that there are no setup fees and no hidden costs. Note that prices are displayed in USD on the pricing page while the Gladia SAS terms are denominated in euros excluding tax. Tiers can be changed at any time from the account settings.
- Starter — pay-as-you-go at USD 0.61/h asynchronous and USD 0.75/h real time
- €50 of credits on signup
- 30 concurrent real-time requests and 25 asynchronous
- automatic language detection and switching
- diarization
- 100+ languages
- GDPR/HIPAA/SOC 2 Type 2
- help center and Discord support
- Growth — from USD 0.20/h asynchronous and USD 0.25/h real time against an upfront commitment
- flexible concurrency
- volume discounts
- automatic training opt-out
- 99.9% uptime SLA
- priority processing queue
- help center and Discord support
- Enterprise — custom pricing on an annual plan
- custom models and fine-tuning
- unbundled pricing
- unlimited concurrency
- training opt-out by default
- zero data retention
- SLA
- premium support with a dedicated Slack channel and account manager
- All three tiers include asynchronous transcription
- real-time streaming
- speaker diarization
- 100+ languages
- word-level timestamps and SOC 2 Type II
Data, GDPR & hosting
A consolidated view of how Gladia handles your data.
GDPR overview
GDPR implementation is concrete and documented. The publisher, GLADIA SAS, is established in Rennes, France, under French law, with a designated DPO reachable at privacy@gladia.io or by post (DPO GLADIA, 90B rue de Fougères, 35700 Rennes). The legal notice cites rights of access, rectification, erasure, objection, restriction and portability under French Law 78-17 as amended and Regulation (EU) 2016/679, with recourse to the CNIL. A DPA is published in full on the site with its annexes, and separate agreements exist for EU and US customers; a BAA is available for healthcare use. The privacy notice is dated 07/07/2025 and the legal notice 13/08/2026. Claimed certifications: GDPR, HIPAA, SOC 2 Type II and ISO 27001. No Article 27 representative is named, and none is required since the main publisher is EU-established. A second entity, Gladia Inc. (Delaware), handles US contracts.
Who owns the data?
Gladia's Compliance Hub states that customer data remains fully owned by the customer, with logical isolation per account and no data shared between customers. The privacy notice splits Gladia's role in two: it acts as a data processor on behalf of its clients for the audio and text they submit, and as a data controller for its own purposes such as account management, marketing and recruitment. Employee access to production data is disabled by default; temporary access requires explicit approval and is logged. Data processed as a processor is deleted on the client's instruction, and deletion can also be requested from support@gladia.io.
Reuse rights
Transcripts and enrichment outputs belong to the customer, who can reuse, redistribute and build on them without asking Gladia for permission. Submitted audio is processed only to deliver the requested service, and the Compliance Hub states that data is isolated by default and never used for model training unless explicitly agreed. The pricing page confirms automatic training opt-out on Growth and opt-out by default on Enterprise, and the home page comparison table lists training opt-out as included "By default" where Deepgram, AssemblyAI and Speechmatics are marked as paid opt-out. One contradiction must be flagged: the /security page, last updated 09/02/2024, still states that only users in the "Free Plan" are subject to data used for model training — a plan name that no longer exists on the current pricing page, and a rule that conflicts with the Compliance Hub. Separately, the privacy notice describes Gladia's training datasets as several thousand hours of human voices across more than a hundred languages, stripped of identifying data, and states that they contain no voice recordings from use of the AI Service, with no commercial exploitation and no individual profiling. Until the site is aligned, have the training rule that applies to your contract confirmed in writing.
Data retention & training
Hosting summary
By default Gladia processes audio on a European cloud provider located in France; other geographies and specific providers are offered on customer request. Annex III of the published DPA names seven subprocessors with their data locations: OVH Cloud (instances and object storage in Gravelines, France) handles the core audio and text processing; Shadow (Paris) provides storage and processing in France; AWS runs Cognito identity in the US, S3 in the EU for Gladia EU and S3 plus CloudWatch in the US for Gladia US; Modal Labs (diarization) operates in the EU and US; Google Cloud (sentiment, translation) in the US and EU; OpenAI Ireland Limited (Audio-to-LLM, translation, summarization) in the US and EU; and DeepL SE (translation, Cologne) in the EU. EU and US workloads run on separate regional infrastructure, the home page advertises 100% EU data residency, and custom hosting is available at Enterprise level. Data is encrypted in transit (TLS) and at rest. One distinction matters: the marketing website itself is hosted by Webflow Inc. and Vercel Inc. in the United States, which is the public site and not the processing infrastructure.
Things to keep in mind
Risks and trade-offs to weigh before adopting Gladia.
- The site contradicts itself on model training: /security, last updated 09/02/2024, still states that "Free Plan" audio feeds model training and that only "Pro" and "Enterprise" are excluded, while the Compliance Hub says training is excluded by default. Get the rule that applies to your contract in writing before uploading sensitive audio
- The same /security page presents ISO 27001 as work in progress while the rest of the site announces it as obtained, and the audit evidence (SOC 2 and HIPAA reports, penetration test, current subprocessor list) is only released on request through the Trust Center, so until you ask you are trusting a claim
- Two separate contracting entities, Gladia SAS and Gladia Inc., with separate terms: check which one binds you before signing. The published US address is a registered agent (Corporation Service Company, Wilmington), not an operational office
- Prices are shown in USD on the pricing page while the Gladia SAS terms are denominated in euros excluding tax, so budget for currency and tax gaps
- The €50 credit is a one-time grant with no monthly reset. It is easy to mistake for a free plan and easy to burn through in a few days of load testing, especially with auto top-up enabled
- Default retention is 12 months and only comes down through configuration or tier. Recorded conversations are among the most sensitive data an organization holds, and leaving the default in place means a year of transcripts sitting on infrastructure you do not control
- Automatic transcription invites over-trust. The product ships hallucination filters and post-processing, which is an acknowledgement that errors happen: speaker labels, medical or legal terms and PII redaction deserve human review rather than blind acceptance
Setup & Integrations
Technical difficulty
Moderate and developer-oriented: any programmatic use requires an API key and code. Gladia claims an average integration time under 4 hours from first call to production traffic, with official Python and Node.js SDKs, a CLI and reference WebSocket clients. Prebuilt integrations exist for Pipecat, LiveKit, Vapi, Retell, Twilio, Recall.ai, Meeting BaaS, Attendee and VideoSDK, plus migration guides from Deepgram and AssemblyAI. Non-developers can test in the Playground or automate through Zapier, Make and n8n. ScorePlay reports going live in under a day of development; Livestorm reports removing the scaling burden of self-hosted Whisper, with a claimed 20% less DevOps effort.
Deployment
Integrations
Supported languages
Behind Gladia
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Gladia.
Frequently asked questions
Can I try Gladia for free?
Is there a permanently free plan?
How many languages does Gladia support?
Does Gladia handle real-time transcription?
Is my audio used to train Gladia's models?
Where is my data hosted?
Which certifications does Gladia claim?
Is a data processing agreement available?
How long is my data kept?
Is there a desktop or mobile app?
Should you pick Gladia?
Gladia is an infrastructure component, not a consumer app. It sells one thing well: a single API that turns audio into text and then into structured meaning — diarization, translation, entities, sentiment, chapters, summaries — across 100 languages, asynchronously or live at sub-300 ms latency.
Its strongest argument is the combination of breadth and jurisdiction. Few competitors match 100 languages with native code-switching, fewer still include enrichment in the base rate instead of charging for it, and almost none process audio by default on a European provider in France with EU and US workloads kept apart. For a European engineering team under GDPR, HIPAA or SOC 2 pressure handling multilingual volume, that mix is hard to reproduce by assembling separate vendors.
The weak points are commercial and documentary rather than technical. Only Starter has a public price; Growth and Enterprise require a sales conversation, so budgeting stays an estimate until you negotiate. And the security documentation has drifted: the /security page has not been revised since February 2024 and still contradicts the newer Compliance Hub on model training, plan names and ISO 27001 status. Nothing suggests bad faith, but confirm in writing which training rule applies to your account before committing sensitive audio.
Who should buy it: technical teams building meeting assistants, voice agents, CCaaS platforms or media tooling, especially those leaving a self-hosted Whisper setup. Who should not: non-technical users wanting a finished application, anyone needing text-to-speech, and teams requiring a permanently free tier — the €50 credit is a one-time grant, not a recurring allowance.
Judged on what it claims and documents, Gladia is a credible, sovereignty-minded audio backbone with a pricing page that stops short and a security page overdue for a refresh.
- Choosing a selection results in a full page refresh.
- Opens in a new window.