Gradium open-sources Phonon, an on-device TTS that starts speaking in milliseconds

What Gradium released

On 20 October 2026 Gradium open-sourced Phonon, a family of small text-to-speech models meant to run entirely on the user's hardware. A day later, at NVIDIA GTC Berlin, the company repeated the launch in a press release that named Hugging Face, Arm, Qualcomm, Appning by FORVIA, LM Studio, Ollama and MLX Audio as partners.

Phonon comes in two sizes. Phonon 31m is an English model with about 31 million parameters; its 8-bit GGUF weights are 64.5 MB. Phonon 40m is a 40-million-parameter line with separate English, French, German, Spanish and Portuguese checkpoints, plus one multilingual checkpoint that covers all five languages. Gradium says the models were announced in private form in April 2026 and sat in a private beta until this open release.

The company is a Kyutai spin-out with offices in Paris and San Francisco and backing from NVIDIA. That lineage matters for the stack: Phonon builds on Kyutai's Pocket TTS work and the Mimi audio codec. Weights for the English 31m model are on Hugging Face; the shared runtime lives in the gradium-ai/xn-ptts repository.

How Phonon is built

Architecturally, Phonon is a causal transformer that takes a text-plus-speaker prefix and predicts one audio latent every 80 milliseconds. A SentencePiece tokenizer with a 4,000-token vocabulary turns the prompt into text embeddings; speaker embeddings condition the voice. A one-step flow-matching head produces the latents, a small Mimi decoder turns them into 24 kHz mono audio, and a separate end-of-speech head decides when to stop.

The 31m model uses six transformer layers, width 512 and eight attention heads. The 40m model uses nine layers at the same width and head count. Gradium says both sizes are distilled from a larger teacher and were trained on NVIDIA H100 clusters with NCCL and CUDA Graphs. The same weights ship with backends for CPU, Metal, CUDA, Vulkan, WebGPU and Apple Core ML, so a phone app and a browser tab can share one model file.

That design is aimed at places where a cloud round-trip is slow, expensive or unavailable: offline assistants, accessibility tools, in-car systems, robots and web apps that want speech without shipping audio to a server. Gradium's demos include Autonomous's Lamp robot speaking in character, a Modal and Pipecat voice bot, an Apple Watch brew guide that keeps working with Wi-Fi off, and a Chrome extension that reads a page locally.

Naturalness and accuracy numbers

Gradium's quality claims rest on two public tests. For naturalness, independent listeners compared Phonon with Kokoro, Magpie, NeuTTS and Pocket TTS in blind pairwise tests across languages (1,189 raters and 17,270 votes). On the English Elo chart in the release post, Phonon 40m scores 2117, Phonon 31m scores 2098 and the multilingual 40m scores 2097, ahead of Pocket TTS 100m at 2056, Kokoro at 2043, NeuTTS at 2034 and Magpie at 1939.

For accuracy, Gradium published a hard set of 500 sentences per language built from phone numbers, emails and other structured text that smaller models often misread. On that English set the reported word-error rates are 1.70% for Phonon 31m, 2.11% for Phonon 40m multilingual, 2.22% for Phonon 40m, 3.50% for Kokoro, 7.40% for Kitten TTS, 8.71% for NeuTTS Nano and 27.51% for Magpie. The set is on Hugging Face as gradium/tts-hard-sentences-202610; transcripts were scored with an automatic pipeline Gradium documents in the post.

Those numbers are Gradium's own evaluations. Blind preference and a hard WER suite are useful signals, but they are not a third-party leaderboard, and Magpie and Kokoro were run with the voices and packages Gradium lists in the footnotes. Teams that care about a specific language or domain still need to measure on their own prompts.

Speed on phones, laptops and small boards

Latency is the other half of the pitch. On a MacBook Air M5 CPU with four threads, Gradium reports Phonon 31m at 29.9 times real time and a first-audio time around 7 ms, against 517 ms for Kokoro on the same sentence (Kokoro does not stream, so its first audio is the whole utterance). With Apple Accelerate and Arm KleidiAI, Gradium's chart reaches 46.7 times real time on that laptop class.

On a Raspberry Pi 5, Phonon 31m is listed at 38 ms to first audio and 3.7 times real time, while Kokoro needs about 7 seconds to start and runs below real time in the same setup. On an NVIDIA L40S the CUDA path is listed at 5 ms to first audio; on an AMD EPYC 7413 CPU it is 43 ms. Arm partners measured 28 ms to first audio with KleidiAI; Qualcomm partners measured 7.5 ms and 24 times real time on Snapdragon 8 Elite Gen 5 hardware.

The practical point is that the same small checkpoint is meant to cover offline phones, browser WebGPU tabs and GPU servers without a separate cloud-only model. WebGPU first-audio times are higher than CPU because the browser path batches frames; Gradium's footnotes spell that out for anyone reproducing the bars.

Licenses, voices and partners

Licensing splits the family. Phonon 31m English is under Apache 2.0, which fits commercial apps that want to ship weights. Phonon 40m language-specific and multilingual checkpoints are under CC BY-NC 4.0, so the larger models are free to study and fine-tune for non-commercial use but need a separate deal for product use. The xn-ptts runtime itself is dual-licensed MIT OR Apache-2.0.

The 31m English card lists 37 built-in voices. The multilingual line advertises 102 built-in voices, including twelve English character voices made with Gradium's voice-design tool. Instant voice cloning and text-to-voice design stay on Gradium's API and Studio; once a voice file is downloaded it can run fully on device inside Phonon.

Steerability at launch is mainly speaking speed and loudness normalization, plus text normalization for numbers and dates. Gradium also says it trains custom Phonon variants with design partners for tighter memory budgets or extra languages such as Korean, Arabic and Japanese. FORVIA and Appning by FORVIA are named as automotive R&D partners who need offline speech in tunnels and across regional languages.

Where it sits among other speech models

Phonon joins a crowded on-device speech wave. AI Lookout has covered Fish Audio's Drama 3 open-weight TTS, which leans on director-style prompts, and cloud-leaning stacks such as Mistral's Voxtral 3 and xAI's Grok Voice 3.0 API. Phonon's bet is smaller: ship a 65 MB Apache checkpoint that starts audio in single-digit milliseconds on a laptop CPU.

The Kyutai connection also ties it to other open robotics and audio work from that lab, including the REKI robot model. For builders, the useful comparison is not only Elo bars but whether Apache 31m is good enough to avoid a hosted TTS bill, and whether CC BY-NC on 40m forces a paid path for the multilingual voices.

Package availability is already broad for an open release day: Rust and Python packages, a browser npm build, Swift and Core ML paths, Docker images and listed integrations with LM Studio, Ollama and MLX Audio. That distribution layer is as important as the scores for teams that want to try the model tonight.

What we could not verify

We did not re-run Gradium's Elo or WER suites, and we did not independently time the MacBook Air M5, Raspberry Pi 5, L40S or Snapdragon benches. Partner figures from Arm and Qualcomm are reported as partner measurements in Gradium's post. We also did not audit training data provenance beyond what the release posts disclose.

Older Gradium posts describe an earlier ~100-million-parameter Phonon with different Seed-TTS scores. Those figures are not the same checkpoints as the 31m and 40m open-weight family announced on 20 October, so we kept them out of the comparison tables above.

What to watch next

Watch whether other labs publish third-party numbers on Gradium's hard-sentence set, whether the Apache 31m model shows up as a default in Ollama and LM Studio voice features, and whether Gradium opens commercial licensing for the CC BY-NC 40m line. Also watch for extra languages beyond the first five, and for whether car and robot partners ship Phonon in products rather than only in demos.