Gradium Lands $100M Backed by Nvidia: The Voice AI API Developers Should Know

Gradium, a Paris-based voice AI startup backed by Nvidia, raised $100M to commercialize its audio-native language model that achieves sub-200ms latency, beating human conversation turn-taking baselines. The company's API and open-source Gradbot framework target developers building real-time voice agents, with benchmarks showing 155ms median time-to-first-audio and 2ms interquartile range under load.

Nvidia doesn’t write a $30M check on a whim. When they extended Gradium’s seed round to $100M last week, the signal wasn’t about European AI hype — it was a compute bet. Voice agents are coming, they will run on GPUs, and Gradium is building the infrastructure layer under them. The funding story is fine. The API is what you should actually read about. The Latency Gap Developers Should Know Most “real-time” voice AI has a dirty secret: it isn’t real-time. The pipeline — send audio to STT, wait for transcript, call the LLM, generate TTS, stream audio — introduces enough lag that users can feel it. The industry calls this the awkward pause. Gradium published benchmark data from Coval’s independent continuous testing https://www.coval.ai/blog/the-future-of-speech-to-speech-ai-inside-gradium-and-kyutai-s-approach-to-full-duplex-conversation/ 750+ runs per model, refreshed every 30 minutes that puts numbers on the gap: | Provider | P50 TTFA | IQR | WER | |---|---|---|---| | Gradium | 155ms | 2ms | 3.3% | | Cartesia Sonic-3 | 188ms | 100ms | — | | ElevenLabs Turbo v2.5 | 264ms | 28ms | 5.2% | | Deepgram Aura-2 | 313ms | 68ms | 6.4% | Human conversation runs at roughly a 200ms turn-taking gap. Gradium is the only provider below that baseline. The median number P50 TTFA is the headline. The IQR tells the production story. At thousands of concurrent sessions, Deepgram’s 68ms IQR means some users hit 381ms — a stutter. Gradium’s 2ms IQR means the 155ms you see in tests is essentially what you get under load. Why They’re Faster: Audio-Native, Not a Text Adapter Most voice AI companies bolt STT and TTS onto a text LLM and call it a voice agent. Gradium runs on a different substrate. The company is commercializing research from its parent lab, Kyutai — a Paris nonprofit co-founded by Neil Zeghidour, who previously did foundational voice AI work at Google Brain, DeepMind, and Meta. Zeghidour and his co-founders built EnCodec, SoundStream, and Moshi — the underlying papers that most modern voice AI systems, including OpenAI’s, are built on top of. Their Audio Language Model ALM uses Soundstream to compress audio into tokens and runs a transformer that predicts audio sequences directly. This matters for two reasons. First, every LLM advance — speculative decoding, RLHF, distillation — transfers directly to the audio model. Second, it enables genuine full-duplex conversation: the model listens and speaks simultaneously, with no voice-activity detection hacks required. “We were just riding for free the LLM wave, but just with audio.” Neil Zeghidour, CEO, Gradium Gradbot: Voice Agent in 50 Lines Gradium also ships Gradbot https://github.com/gradium-ai/gradbot , an open-source framework for prototyping voice agents. The architecture is a Rust multiplexing engine that coordinates three concurrent streams — STT, LLM inference, and TTS — while managing conversational state, interruptions, and turn-taking automatically. Key details for developers: - Works with any OpenAI-compatible LLM Claude, GPT-4o, Groq, Ollama, LM Studio - Async tool calling with three tracked states: new, pending, answered — gracefully handles API lag - Five languages: English, French, German, Spanish, Portuguese - Install: pip install gradbot The team is explicit about Gradbot’s scope: it’s for prototyping and experimentation. For production, they recommend LiveKit or Pipecat as the orchestration layer, with Gradium handling the STT and TTS. Access and Pricing The API runs over WebSocket with a single key covering both STT and TTS. The free tier gives you roughly one hour of TTS with three concurrent requests — genuinely enough to evaluate the latency claims before paying. Paid plans start at $13/month XS and scale to $340/month M and $1,615/month L . Gradium also runs a startup grant program: over $2,000 in API credits and six months on the M plan, with direct engineering support. Details at gradium.ai/pricing https://gradium.ai/pricing . The One Real Gap Gradium supports five languages. ElevenLabs https://elevenlabs.io supports 32. If you are building a multilingual product, that gap is decisive and Gradium is not the right choice today. For English-first or European-market voice agents, the latency and consistency advantage is significant. For everything else, check back in six months — they publish model updates roughly monthly. The Bottom Line Voice is the interface layer that comes after chat. Agentic systems that can call tools while speaking, handle interruptions, and respond in under 200ms are what the next generation of AI products will run on. Gradium’s bet is that owning the audio infrastructure — not the LLM on top — is where the durable margin lives. Nvidia’s check suggests they’re not the only ones who believe it. Start with pip install gradbot and the Gradium API docs https://docs.gradium.ai/ . The free tier is enough to see whether the latency numbers hold up in your environment.