
Gradium
Gradium is a French developer platform of real-time voice AI models: expressive text-to-speech, streaming speech-to-text, live speech-to-speech translation and voice cloning, all delivered through a single API with optional EU or US data residency and zero data retention.
What is Gradium?
Gradium is voice infrastructure sold to developers. The company builds its own foundation models for speech and exposes them through a single API rather than shipping a finished application. Three model families sit at the centre: text-to-speech, billed at one credit per character, speech-to-text at three credits per second, and speech-to-speech translation at thirty credits per second, with a speech-to-text translation variant at four. A separate on-device model, Phonon, runs text-to-speech offline on edge hardware.
Access is through WebSocket streaming or one-shot REST calls, authenticated with an API key. A published OpenAPI 3.1.0 specification describes ten endpoints on api.gradium.ai, covering synthesis, recognition, speech-to-speech, voice management, pronunciation dictionaries and credit usage. A Python SDK wraps them. Alongside the API, Gradium Studio gives a browser playground for voices, pronunciation rules and privacy settings, while Gradbot, an open-source framework published on GitHub, handles the event loop of a voice agent: voice activity detection, turn-taking, interruptions and fillers, on top of any OpenAI-compatible language model.
Two controls matter to enterprise buyers. Data residency pins a session to European or American servers and refuses the request outright rather than serving it from the wrong zone, reporting the zone it used on every response. Zero Data Retention stops request and response payloads from ever being written to disk. Teams that need more can run inference inside their own AWS account through SageMaker, on Baseten, or on their own premises.
The company behind it is young. Gradium is a French simplified joint-stock company registered in Paris on 16 July 2025, founded by Neil Zeghidour, Laurent Mazaré, Olivier Teboul and Alexandre Défossez, the researchers behind the Kyutai laboratory. In July 2026 it announced that its seed financing had been extended to one hundred million dollars, with NVIDIA among the new investors, and opened an office in the San Francisco Bay Area. Five languages are supported: English, French, Spanish, Portuguese and German.
What it does
- Turn text into expressive, low-latency speech, streamed word by word
- Transcribe live audio into text, with turn detection based on meaning rather than silence
- Translate speech into speech or into text in real time across five languages
- Clone a voice from ten seconds of audio, or in high fidelity from thirty minutes
- Run a text-to-speech model offline, on the device, with no network call
- Assemble a production voice agent in about fifty lines with the open-source Gradbot framework
- Pin every session to the EU or to the US and prove it from the response headers
When to use Gradium / When not to
A quick filter to help you decide if Gradium is the right fit.
When to use Gradium
- Backend and full-stack developers wiring voice into an existing product or agent pipeline
- Startup CTOs and founders who need a public price list and a permanent free tier before committing
- Enterprise teams whose security review demands EU or US data residency, zero data retention, or inference inside their own VPC
- Localisation and translation providers adding live speech-to-speech translation across English, French, Spanish, Portuguese and German
- Embedded and mobile engineers who need text-to-speech running offline on the device
When not to use Gradium
- Consumers looking for a ready-made app: the terms restrict the service to professional and business use, and there is no mobile app
- Anyone building clinical tools: the terms forbid medical advice, diagnosis, assessment and prescription, as well as emergency healthcare
- Teams whose users speak outside English, French, Spanish, Portuguese and German, since no other language is supported yet
- Projects that need uninterrupted audio longer than five minutes, because a single session is capped at 300 seconds
- Buyers who require a certified security posture today: no ISO, SOC or equivalent certificate is published anywhere on the site
How to use Gradium
A typical end-to-end flow, from setup to results.
- Try the models with no account on the public agent demo, or on the translation and text-to-speech pages
- Create an account through the sign-in portal, then pick a plan in Gradium Studio
- Start on the permanent free tier, which grants 45,000 credits a month but excludes commercial use
- Explore voices, pronunciation dictionaries and rewrite rules inside Studio before writing any code
- Install the Python SDK and copy the API key from your account
- Call the WebSocket endpoint for streaming, or the REST endpoint for one-shot requests
- For an agent, start from Gradbot: define a voice, a session configuration and instructions, then run the loop
- Connect the pipeline to LiveKit, Pipecat, Vapi or OpenClaw if you already use one of those frameworks
- Ask support to enrol your organisation in EU or US residency, then switch to the regional base URL
- Turn on Zero Data Retention yourself in Studio, under your organisation's privacy settings
Pros & Cons
Pros
- Models built in-house by the team behind much of the published work on neural audio codecs and audio language models
- Fully public pricing across six tiers, with the credit-to-hours conversion spelled out for each model
- A permanent free tier of 45,000 credits a month, with no card and no time limit
- Data residency that fails closed: a session that cannot run in its zone is refused, not silently moved
- Zero Data Retention available in self-service on paid plans, and a complete Data Processing Agreement published in the open
- Deployment inside the customer's own infrastructure through SageMaker, Baseten or an on-premises install
- Published OpenAPI specification, Python SDK, an open-source agent framework and migration guides from four competitors
Cons
- Only five languages are supported, which rules out most of the world's users
- A session is capped at 300 seconds, so long-form content has to be split by the developer
- The free tier forbids commercial use, making the 13 USD tier the real entry point for any product
- Plan credits expire at the end of the month and do not roll over; only pay-as-you-go credits do
- Pro voice cloning starts on the 340 USD tier, well above the entry plans
- No security certification is published: the trust centre is empty to the public and the terms describe further measures as still to come
- The Data Processing Agreement promises a sub-processor list on the website, and no such list exists
Pricing & Plans
A permanent free plan is available at no cost, offering 45,000 credits per month with access to both Studio and the API, but excluding commercial use. The cheapest paid tier, XS, is priced at 13.00 USD per month and is the first plan to permit commercial use. Prices are displayed monthly by default; an annual billing option is offered. Additional credits are sold in blocks of 100,000 at rates falling from 6.90 to 3.80 USD as the tier rises, and the Enterprise tier is quoted on request.
- Free — 0 USD per month — 45
- 000 credits
- Studio and API access
- 5 instant voice clones
- no commercial use
- XS — 13 USD per month — 225
- 000 credits
- commercial use
- 1
- 000 instant voice clones
- extra credits at 6.90 USD per 100
- 000
- S — 43 USD per month — 900
- 000 credits
- extra credits at 5.00 USD per 100
- 000
- M — 340 USD per month — 9 million credits
- 5 Pro voice clones
- extra credits at 4.00 USD per 100
- 000
- L — 1
- 615 USD per month — 45 million credits
- 20 Pro voice clones
- extra credits at 3.80 USD per 100
- 000
- Enterprise — priced on request — unlimited credits and custom concurrency
Data, GDPR & hosting
A consolidated view of how Gradium handles your data.
GDPR overview
Implementation is concrete rather than declarative. Gradium is established in France, so the GDPR applies directly and the CNIL is the supervisory authority. The privacy policy, dated July 2026 and downloadable as a PDF, lists the data collected, the purpose and the legal basis for each processing operation, and the full set of data subject rights, including an explicit right to object to model improvement. Requests go to dpo@gradium.ai or to the Paris office. A complete Data Processing Agreement is published as Exhibit 1 to the terms: sub-processing, international transfers under the 2021/914 standard contractual clauses, the EU-US Data Privacy Framework, the UK Addendum, audit rights, and deletion or return within forty-five days. Gradium commits to assist with Article 32 obligations. No Article 27 representative is named, and none is needed for an EU-established company.
Who owns the data?
The client keeps every intellectual property right in the data it submits. Gradium receives a worldwide, non-exclusive, royalty-free licence, for the length of the subscription, to use, store, copy, modify and reproduce that data in order to run, maintain and optimise the service, and it may pass it to its own service providers. A second, separate licence covers service improvement and AI model training; that one falls away when Zero Data Retention is switched on. Gradium owns the models, the API and the documentation, and any feedback a client sends becomes Gradium's exclusive property, model training included. On personal data, Gradium acts as processor for client content and as controller for account data.
Reuse rights
Outputs belong to the client's side of the contract: the client owns the intellectual property rights in what it submits, and the terms place no restriction on reusing the generated audio or transcripts in its own product, provided the underlying use is lawful. Permission is not required for that reuse, but three duties travel with it. The client must obtain consent from every individual whose voice is recorded, cloned or processed, and must tell people they are talking to an AI system. It must not direct the service to impersonate anyone without authorisation, to produce disinformation or political content, or to give medical, legal or financial advice without the relevant qualifications. Reverse-engineering the models, or using outputs to train a competing model, is expressly forbidden. On the free plan the licence is limited to internal, non-commercial use.
Data retention & training
Hosting summary
By default, requests to api.gradium.ai are routed to the nearest server. Two regional endpoints exist, one for the European Union and one for the United States. Without enrolment, calling a regional endpoint is a best-effort request with no commitment attached. Once an organisation is enrolled, the call carries a guarantee: the session runs inside its zone, or it is refused, with a 403 over REST and a 1008 close over WebSocket. Every response names the zone it used, in a response header over REST and in the ready message over WebSocket, so the customer can assert on it rather than trust a hostname. Enrolment is arranged by the vendor on paid plans, in either an explicit per-call mode or an implicit organisation-wide mode. Data that is not session data — account records, custom voices, pronunciation dictionaries — lives in a control plane hosted in the European Union for every customer, and that holds even under US residency. The marketing site itself is served from Paris, on French infrastructure. The jurisdiction of the contract is France, with exclusive jurisdiction to the Paris civil courts.
Where Gradium works
Country-level availability.
Not available in
Things to keep in mind
Risks and trade-offs to weigh before adopting Gradium.
- Training on customer data is the default: it stops only if you enable Zero Data Retention, which the terms say may cost extra, or if you exercise the right to object
- Voice cloning shifts the consent burden onto you: the terms require explicit consent from every person whose voice is recorded, cloned or processed, and you are the one who must obtain and keep it
- Synthetic speech that is indistinguishable from a real person invites impersonation and fraud; the terms forbid it, but nothing technical prevents it
- The legal entity was registered in Paris under a different name, Sphère II, with a holding purpose, and renamed Gradium days later; the site meanwhile claims a September 2025 founding, two months after the registration
- The domain gradium.ai is registered to Kyutai, not to Gradium itself, so the registrant of the domain and the publisher of the service are two different organisations
- No security certification is published and the public trust centre returns an empty page, so any claim about the security posture rests on the vendor's own description
- The company names healthcare among its customer verticals while the terms forbid medical advice, diagnosis, assessment and prescription; anyone in that sector must reconcile the two before deploying
Setup & Integrations
Technical difficulty
This is a developer product and it assumes a developer. Nothing needs to be installed to evaluate it: the browser demo and Studio playground work straight away. Going further means installing the Python SDK, handling an API key and writing against a WebSocket or REST endpoint, which is routine work for a backend engineer. The vendor claims a first agent in about fifty lines with Gradbot, and migration guides exist from four competing APIs. Switching region is a base URL change, nothing more. Advanced deployment — inference in your own AWS account, Baseten, or on-premises — requires an infrastructure team.
Deployment
Integrations
Supported languages
Behind Gradium
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Gradium.
Frequently asked questions
Which languages does Gradium support?
Is there a free plan, and can I use it in a commercial product?
How long can a single session run?
Can I clone my own voice, and what does it take?
Where is my data processed and stored?
Does Gradium train its models on my data?
What happens when I use up my plan credits?
Can Gradium run inside my own infrastructure?
Is there an API, and how mature is it?
Are there any offers for early-stage companies?
Should you pick Gradium?
Gradium is a serious piece of voice infrastructure from a team with an unusually deep research pedigree, and it behaves like one. The pricing is fully public across six tiers, the credit-to-hours conversion is spelled out rather than hidden, the API is documented with a real OpenAPI specification, and the enterprise controls that usually take three calls to establish — data residency, zero retention, deployment inside the customer's own network — are described in writing before anyone asks.
The reservations are equally clear. Five languages is a narrow footprint. A session ceiling of 300 seconds forces developers to handle chunking themselves. The free tier is genuinely permanent but bars commercial use, so the real entry price is 13 USD a month. No security certification is published anywhere: the public trust centre is empty, and the terms describe further organisational and technical measures as still to come, which is honest but leaves a procurement gap. The promised sub-processor list is announced in the Data Processing Agreement and does not exist on the site.
Training on customer data is the default behaviour, not the exception; switching it off means enabling Zero Data Retention, which the terms say may carry an extra fee. That is worth reading twice before signing.
For a developer or an engineering team building voice agents in one of the five supported languages, particularly one that has to answer a question about where inference runs, Gradium is a credible and well-documented option. For anyone needing broad language coverage, a certified security posture today, or a finished application rather than an API, it is too early.
- Choosing a selection results in a full page refresh.
- Opens in a new window.