Trained on NVIDIA GPUs, Phonon generates natural speech locally in as little as 64.5 MB, on phones, laptops, embedded hardware and in the browser.
BERLIN, NVIDIA GTC, October 21, 2026. Gradium, the voice AI company founded by the cofounders of Kyutai research lab, today released Phonon, the best performing on-device text-to-speech system. Phonon runs entirely on the user's hardware: speech is generated locally, with no network round-trip and no cloud dependency, so text and audio stay on the device.
Phonon has 31m parameters, against 82m for Kokoro, the most widely used open on-device TTS model. On blind human evaluations, Gradium Phonon wins on overall preference, ahead of Kokoro, Magpie and Pocket TTS. Gradium trained the different Phonon variants on a cluster of NVIDIA H100 GPUs, leveraging NCCL for distributed training.
Built for phones, laptops, embedded hardware and the web #
Phonon targets developers who need speech output where cloud calls are slow, costly or unavailable: mobile apps, offline assistants, accessibility tools, in-car and embedded systems, and web applications.
Model sizes. Two variants at launch, one with 40 million parameters and one with 31 million parameters. These small models are sized to ship inside a mobile app or load in a browser tab.
Languages. English, French, German, Spanish and Portuguese in a single multilingual model at launch, with more languages on the roadmap.
Broad hardware coverage. Phonon comes with a custom inference codebase with faster inference than Kokoro and llama.cpp-based runtimes on the same hardware. It supports multiple backends to give the best inference speed on any possible hardware.
Voices. A catalog of ready-to-use voices, including Harper, Sterling, Freya and Toby, plus voice cloning and voice design to create custom voices.
Steerability. Control over speaking speed.
Phonon ships with integrations for Rust, Python, Swift and the browser through WebAssembly.
The Phonon inference stack includes a NVIDIA CUDA backend that allows Phonon to run on professional and gaming GPUs as well as on embedded hardware with the NVIDIA Jetson platform. This gives the opportunity for developers in the gaming and robotics industry to generate natural speech close to their users in the most efficient way. The WebGPU backend allows for running the model in a web browser and benefiting from the acceleration of NVIDIA GPUs.
On desktops and servers with NVIDIA GPUs, Phonon runs inference through its CUDA backend 8x faster than on CPU, generating the first bit of speech in 5ms on an NVIDIA L40S against 43ms on a AMD EPYC 7413. Developers get the same model on both: CPU for phones and laptops, NVIDIA GPUs for high-throughput serving.
Quotes #
"Voice is becoming the main way we interact with devices and software. Only on-device models can guarantee availability, privacy and cost for this level of daily use," said Laurent Mazaré, CTO of Gradium. "Phonon puts natural, controllable speech on a phone or in a browser tab at 90 MB. Training it with CUDA Graphs on NVIDIA GPUs let our team iterate fast on a very small model and push its quality past larger open models. We are releasing it as open source so every developer can build with it, and connect to the Gradium API when they need voice cloning or voice design."
Availability #
Phonon is available today. The multilingual model weights are on Hugging Face at https://huggingface.co/gradium/phonon-31m-en and the inference code is on GitHub at github.com/gradium-ai/xn-ptts, under the Apache 2.0 licence. Voice cloning and voice design are available through the Gradium API at gradium.ai. Launch partners include Hugging Face, Arm, Qualcomm, Appning by FORVIA, LM Studio, Ollama, MLX Audio.
About Gradium #
Gradium builds voice AI models for real-time voice interactions: text-to-speech, speech-to-text, voice cloning, live translation and on-device models. Spun out of the Kyutai research lab and backed by NVIDIA, Gradium has offices in Paris and San Francisco. Learn more at gradium.ai.
Media contact: contact@gradium.ai
Related posts #
Oct 20, 2026AnnouncementDiscover Open-Sourcing Phonon: A Suite of Frontier On-Device TTS Models
Today, we're open-sourcing Phonon, now a family of 31m and 40m parameter text-to-speech (TTS) models that set a new standard for naturalness and accuracy on-device.
Jul 15, 2026ResearchDiscover Smaller, clearer, better: Phonon Multilingual beats Magpie and NeuTTS Nano in French, German, and Spanish
Phonon, our 100M-parameter on-device Text-To-Speech model, produces up to 3.5x fewer word errors than NVIDIA's Magpie TTS in French, German, and Spanish at 3.6x fewer parameters, and leads NeuTTS Nano on both word error rate and speaker similarity. It clones a reference voice across five languages: English, French, German, Spanish, and Portuguese.
Apr 3, 2026AnnouncementDiscover Gradium Phonon: On-Device TTS for Consumer Apps, NPCs, and Offline Products
Announcing Gradium Phonon, our new on-device text-to-speech model designed for consumer apps, NPCs, and offline products.