Summarize this article with:
Gradium brings real-time voice AI models built in France to Eden AI: first audio in as little as 70 ms in our tests, 1,400 voices in five languages, one API.
In a voice product, latency is not a metric on a dashboard. It is the pause your user hears before the answer starts, and every extra fraction of a second makes an agent sound less like a conversation and more like a machine.
Gradium is built around removing that pause, and it is now available on Eden AI.
Who Gradium is
Gradium is a voice AI company based in Paris. It was founded in September 2025 by Neil Zeghidour, Laurent Mazaré, Olivier Teboul and Alexandre Défossez, the researchers behind Kyutai, the AI research lab known for pioneering real-time speech systems.
Its focus is narrow and deliberate: real-time voice for voice agents, with text-to-speech and speech-to-text models designed for live conversation first.
Speed: as low as 70 ms to the first sound
We measured it ourselves. On 19 October 2026, from France, on Gradium's EU streaming API, the first chunk of audio arrived as fast as 70 milliseconds after the text was sent, network round trip included, across 32 requests over WebSocket and REST.
That number matters because of where it sits. In a voice agent, the model is only one link in the chain: speech recognition, the language model, network hops and, on the phone, the carrier itself all add their own delay. When the voice model's share drops to a few tens of milliseconds, most of the budget is left for the rest of the pipeline, and the conversation feels natural instead of stilted.
Gradium also makes a point that is easy to miss: speed usually comes at the cost of accuracy, and its model was built to be both fast and robust rather than trading one for the other.
Built in France, with EU processing
Gradium's models are built in France. On Eden AI, Gradium's routes are in the EU region, and according to Gradium's documentation, organisations enrolled in its EU data residency get sessions that are processed in the EU and do not leave it.
1,400 voices, and a simple price
Gradium's catalogue counts around 1,400 voices covering 28 accents across English, French, Spanish, German and Portuguese, cast for customer service, narration, ads and characters. For teams building for European users, native French, Spanish or German voices are not an afterthought.
On Eden AI, Gradium text-to-speech costs $57.80 per million characters, so you know the cost of a script before you generate it.
Try it through Eden AI
Gradium sits behind the same Eden AI key as every other provider, and its models are listed on the Gradium provider page. A text-to-speech call is one request, and the response gives you the generated audio file:
curl https://api.edenai.run/v3/universal-ai \
-H "Authorization: Bearer $EDENAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "audio/tts/gradium",
"input": {"text": "Bonjour ! Your order has shipped and will arrive tomorrow."}
}'Speech-to-text works the same way with audio/speech_to_text_async/gradium, and audio/speech_to_text_async/gradium/stt-translate transcribes and translates in one step, at $0.0104 and $0.0139 per minute. On Eden AI these speech-to-text routes are asynchronous: the _async describes our route, not Gradium's model, which streams in real time on Gradium's own API.
Swap gradium for another provider in the model string and you can compare voices, price and output on the same text, with the same key and the same bill.
Why we are glad to have Gradium on board
Voice is where the gap between a demo and a product is the most audible. A team can build a convincing prototype with any text-to-speech engine, then discover in production that a noticeable pause before every answer is enough to make the conversation feel broken.
Gradium tackles exactly that problem, with a team that comes straight from real-time speech research. For developers who need a voice that keeps up with the conversation, it is now one model string away.
Where to start
Pick the part of your product where users wait for a voice, an IVR greeting, a support agent's first sentence or a narration preview, and generate it with Gradium through Eden AI. Compare it with what you run today on the same script: the voice, the price per call, and how it sounds to someone who is waiting.
