Gradium builds voice AI models and developer infrastructure for companies that want fast, natural speech interfaces. The Paris startup has now pushed its seed financing to $100 million by adding new investors including Nvidia, a sharp sign that the Gradium voice AI bet is bigger than a niche European lab spinout. Voice agents still break the illusion the second latency creeps in or pronunciation falls apart on things like phone numbers, codes, and email addresses. Founded in September 2025 by Neil Zeghidour, Laurent Mazaré, Olivier Teboul, and Alexandre Défossez, Gradium is trying to fix that with real-time speech infrastructure built for production, not just flashy demos.
What is Gradium voice AI and how does it work?
At a basic level, Gradium sells the plumbing for voice apps. Developers can use its APIs for text-to-speech, speech-to-text, speech-to-speech translation, voice cloning, and on-device text-to-speech. It supports REST for one-shot jobs or WebSockets for live conversations. That matters.
For a customer building a voice agent, the workflow is pretty direct. Audio can be streamed in over WebSocket. Gradium transcribes it in real time, and its speech stack sends back transcripts, turn-taking signals, or synthesized speech as they’re produced. In speech-to-speech mode, the system handles transcription and translation. It also re-synthesizes over a single duplex connection, so the developer doesn’t have to stitch separate tools together by hand.
The feature list is more practical than glamorous. Gradium’s API supports 5 languages today — English, French, German, Spanish, and Portuguese — with voice selection across those languages. It also offers voice cloning from a sample as short as 10 seconds, semantic voice activity detection messages every 80 milliseconds, and adaptive delay controls so builders can trade a bit of latency for accuracy when they need to. That’s what enterprise buyers care about.
Gradium is already widening the stack. It has launched GradBot, an open-source framework that helps developers prototype voice agents in around 50 lines of code. It also launched Phonon, an on-device TTS product that runs fully on CPU across iOS and Android for offline, privacy-sensitive, or high-volume use cases. Phonon is in private beta, while the main cloud platform is already live.
Who founded Gradium and why are investors betting big?
The founding story
Gradium was founded in September 2025 and emerged from stealth in December 2025 as the first startup spun out of Kyutai, the French AI lab backed by Xavier Niel. The idea was simple enough: take frontier speech research out of the lab and turn it into infrastructure developers can actually ship. By July 8, 2026, the company had already reopened its seed round and expanded it to $100 million.
Why this team looks unusually credible
This isn’t a first-time founder team learning speech tech on the fly. Neil Zeghidour, Gradium’s CEO, previously worked at Meta and Google DeepMind, while Olivier Teboul came from Google Brain. Laurent Mazaré worked at Google DeepMind and Jane Street. Alexandre Défossez came from Meta.
Gradium says these founders helped shape core methods behind modern audio language models. It’s an ambitious pitch, but it fits the team’s research pedigree.
Zeghidour is also the connective tissue between Kyutai’s research culture and Gradium’s commercial push. Researchers behind Kyutai’s real-time speech work built the company, and that ongoing relationship gives Gradium a cleaner pipeline from research to product than most startups get. That matters in voice. Model quality, latency, and robustness still move fast enough that last year’s edge can disappear quickly.
Early traction, fundraising, and what the round says
The fundraising history is already chunky. Gradium launched with a $70 million seed in December 2025 from FirstMark Capital, Eurazeo, DST Global, Eric Schmidt, Xavier Niel, Rodolphe Saadé, Korelya, and Amplify Partners, then extended that round to $100 million with new investors including Nvidia. The company generated first revenue within weeks of launch and has early adopters across gaming, AI agents, customer care, language learning, healthcare, and education. The source article adds one named customer: Renault.
How does Gradium compare with rivals?
This is a crowded fight. ElevenLabs is the obvious private-market benchmark after raising $500 million at an $11 billion valuation in February 2026, and Google is a serious threat wherever Gemini’s real-time voice stack shows up. Gradium’s own research posts also benchmark against GPT real-time translation, Gemini live translation, Cartesia, Inworld, and multiple ElevenLabs voice models. That tells you where it thinks the battle is.
Its pitch is narrower and more technical than the generic “we do voice” line. Gradium leans hard into ultra-low latency and semantic turn detection. It also emphasizes pronunciation accuracy on enterprise content and a developer stack that covers cloud APIs, voice cloning, translation, and offline deployment. The real incumbent alternative is still a stitched-together stack of separate speech recognition, LLM, and text-to-speech vendors. That kind of setup creates handoff lag and brittle conversations. Gradium is betting companies would rather buy a tighter stack from one provider if the performance is there.
Why did Gradium voice AI reopen its seed round?
The money matters because it changes the scale of the company’s ambition. A $100 million seed round is already unusual. Reopening it only 7 months after launch, then bringing Nvidia onto the cap table, makes it look less like a normal early-stage extension and more like a fast acceleration round without the label.
The Bay Area move is just as telling. The new capital will fund AI research, product work, international expansion, and a new San Francisco Bay Area office. That isn’t subtle. It’s a talent war move, and it’s also a customer proximity move — closer to the labs, infra partners, and enterprise builders shaping the current voice-agent market. For customers, it may mean faster product shipping and more direct support in the US. For Nvidia, it’s a bet that real-time voice workloads will keep eating GPU and edge compute.
How big is the voice AI market in 2026?
The broader market is already large enough to justify this kind of spending. One recent forecast pegs the global speech and voice recognition market at $9.66 billion in 2025 and projects it will reach $23.11 billion by 2030, which implies a 19.1% compound annual growth rate. That’s not a tiny experimental category anymore. It’s a fast-growing infrastructure market.
The demand shift isn’t just about better synthetic voices. Buyers now want real-time voice agents and multilingual support. They also want lower latency and more flexible deployment models, including offline or privacy-sensitive setups where cloud-only inference won’t work. That’s why Gradium’s mix of streaming APIs, voice cloning, live translation, and on-device TTS makes sense right now. The market is moving from “can the model speak?” to “can the product actually survive production?”
What’s next for Gradium voice AI?
The interesting part isn’t just that Gradium raised more money. A Paris team spun out of Kyutai convinced Nvidia to join a seed round extension less than a year after founding the company, then used that momentum to plant a flag in the Bay Area. Gradium voice AI still has to prove it can turn research firepower into durable developer adoption. The open question is whether it becomes core infrastructure for voice agents, or just another well-funded model vendor in a brutal category.
Read how SAVE Microfinance raised ₹40 Cr in debt funding from Indian Overseas Bank and Northern Arc Capital to expand collateral-free rural lending for underserved women entrepreneurs across India.
FAQ
- What funding did Gradium raise in 2026?
Gradium extended its seed round to $100 million on July 8, 2026, and added new investors including Nvidia. The company had previously launched out of stealth in December 2025 with a $70 million seed backed by FirstMark Capital, Eurazeo, DST Global, Eric Schmidt, Xavier Niel, and others. - How does Gradium’s product actually work for developers?
Gradium gives developers APIs for speech-to-text, text-to-speech, live speech translation, and voice cloning, with REST for batch use and WebSockets for live conversations. A builder can stream text or audio in and get transcripts or synthesized speech back in real time. They can also clone a voice from a 10-second sample or deploy an offline TTS model through Phonon. - Who founded Gradium?
Gradium was founded in September 2025 by Neil Zeghidour, Laurent Mazaré, Olivier Teboul, and Alexandre Défossez. The team came out of Kyutai and brings experience from Meta, Google DeepMind, Google Brain, and Jane Street. That helps explain why investors were willing to write unusually large checks this early. - What market is Gradium competing in?
Gradium is competing in the voice AI infrastructure market, especially around real-time voice agents, speech APIs, and multilingual speech systems. That puts it up against specialist platforms like ElevenLabs as well as bigger model companies such as Google. It also competes with the older do-it-yourself approach of stitching together separate speech, language, and synthesis tools.
