Phonon: A 64 MB TTS Model Whose Three Spec Sheets Don't Agree

The English checkpoint of Phonon 31M — published as gradium/phonon-31m-en — is a 64.5 MB file. That is the whole model, quantized to eight bits, and it produces 24 kHz speech on a laptop CPU with a first-audio figure the vendor puts in single-digit milliseconds. Gradium, a Paris speech company founded by former Kyutai researchers, put the Phonon family on Hugging Face on 20 October 2026, alongside a Rust runtime and a WebAssembly build that lets the thing speak inside a browser tab.

Three days later, the most useful thing a reader can do is not listen to the demo. It is to put the release blog, the model card and the runtime README next to each other. They describe the same two checkpoints — Phonon 31M and Phonon 40M — and they do not agree on the numbers. That disagreement is not a scandal; it is normal for a three-day-old release. But it tells you exactly which figures are safe to plan around and which are still marketing in transit.

What actually shipped on 20 October

Gradium released four things at once: weights, voices, a runtime, and a licence split that matters more than the benchmark table.

The weights come in three shapes. Phonon 31M is English-only, Apache 2.0, 37 built-in voices, and a q8 GGUF of 64.5 MB. Phonon 40M ships as five separate per-language checkpoints — English, French, German, Spanish and Portuguese — each around 74.6 MB, each gated behind a Hugging Face terms acceptance, and each under CC BY-NC 4.0. A third checkpoint, Phonon 40M multilingual, folds all five languages into one model with 102 voices, under the same non-commercial licence.

That licence line is the first decision point, and it is easy to miss because the family is described as open. Only the 31M English model is free for commercial use. The 40M models are open weights in the research sense: downloadable, inspectable, runnable, but not licensed for a product without a commercial agreement. A team that prototypes on the 40M because it sounds better and then tries to ship will hit that wall late and expensively.

The architecture, in one paragraph

Phonon is a flow-matching transformer that predicts audio latents in 80 millisecond frames and hands them to Kyutai's Mimi codec for decoding. It builds on Pocket TTS, the small-model line that came out of the same research circle. The runtime, xn-ptts, is written in Rust and does not use PyTorch at inference: CPU paths use Arm KleidiAI, Apple devices can use Core ML, Snapdragon chips can use the QNN NPU, and browsers get a SIMD WebAssembly build. There are bindings for Python, Swift, JavaScript and C, plus a Docker image that exposes an OpenAI-compatible speech endpoint.

The practical consequence is that Phonon is not a checkpoint you load into a Python notebook and hope fits. It is a deployment target. The same 64 MB file runs on a Raspberry Pi, an iPhone, a browser and a container, and the runtime is the product as much as the weights are.

Three documents, three sets of numbers

Here is the part worth slowing down for. On the 20 October release blog, Phonon 31M reaches first audio in 7 milliseconds on a MacBook Air M5 CPU and runs at 11.6 times real time on one core, 29.9 times on four threads. On a Raspberry Pi 5 it starts in 38 ms and runs at 3.7 times real time. The same blog gives Phonon 31M a hard-text word error rate of 1.70% and an English naturalness Elo of 2,098, against Kokoro's 2,043.

The runtime README, updated in the same window, gives the same MacBook Air M5 a throughput of 26 times real time at 5 ms time-to-first-byte on one core, and 41 times at 6 ms on four cores. On the Raspberry Pi 5 it says 2.3 times real time with 68 ms to first audio. Those are not rounding differences. One document says the model is roughly twice as fast on a Pi as the other does.

The Hugging Face model card adds a third version: 1.5% hard-set WER and 0.51% on Seed-TTS for Phonon 31M, with the same 2,098 Elo. The release blog's table says 1.70% on the hard set. Both are vendor-run evaluations. Neither cites an independent harness, and there is no Artificial Analysis entry for the family yet.

• Parameter count — 31M excluding the audio tokenizer, but Hugging Face metadata lists 41.1M for the same checkpoint.

• First audio — 7 ms on the release blog, 5 ms in the README, on the same MacBook Air M5.

• Throughput on one core — 11.6× real time in the blog, 26× in the README.

• Raspberry Pi 5 — 38 ms and 3.7× in the blog, 68 ms and 2.3× in the README.

• Hard-set WER — 1.70% on the blog, 1.5% on the model card.

This is not a reason to discard the release. It is a reason to treat every number as a range until someone outside Gradium reproduces it. The Elo figures are the most comparable because they name their baseline: Phonon 40M at 2,117, Phonon 31M at 2,098, Phonon 40M multilingual at 2,097, Kokoro at 2,043. A 55-point gap over Kokoro on the vendor's own listening test is a meaningful claim, and it is falsifiable, which is more than most TTS launches offer.

What the repo tells you that the blog doesn't

The repository answers questions the announcement skips. The browser package can load only Phonon 31M directly from Hugging Face; the 40M checkpoints have to be served by your own application, which is a real constraint if the pitch that caught your attention was "TTS in a tab." Gated models need a Hugging Face token at build or run time. The Android QNN path requires Android 12 and a supported Snapdragon HTP, and it has no automatic CPU fallback — if the NPU rejects the model, you get an error, not a slower run.

Voice files are separate from weights. The checkpoints ship some voices, and a catalogue at gradium/phonon-voices carries more, including custom voices in .safetensors format. That separation is sensible engineering and also the part with the least documentation. How a custom voice is trained, what consent record travels with it, and whether Gradium watermarks generated audio are all unanswered in the public material as of 23 October.

What is confirmed: the code is MIT or Apache-2.0 at your option, the 31M weights are Apache 2.0, the 40M weights are CC BY-NC 4.0, and the runtimes genuinely exist for Linux, macOS, Windows, iOS, Android, the browser and Docker. What is not confirmed: independent latency, independent intelligibility, long-form stability beyond a few sentences, and how the multilingual checkpoint handles code-switching inside a sentence.

Where 31 million parameters is the right tool

The models Phonon is being compared to in the first week are mostly cloud systems with a hundred times the parameter count and a network hop in front of them. That is a category difference, not a quality difference. A 64 MB English model that starts speaking in single-digit milliseconds is for the cases where the audio must not leave the device, or must work when the network does not: offline accessibility tools, game dialogue, kiosk and appliance interfaces, private assistants on a phone, and browser features that cannot afford a round trip.

It is a poor fit for the opposite cases. If you need thirty languages, 48 kHz output, voice design from a text description, or studio-grade cloning from a ten-second reference, Phonon 31M does none of those. Even Phonon 40M multilingual covers five languages, not thirty, and its licence rules out most commercial products.

The honest summary after three days is that Gradium has shipped an unusually complete on-device stack with unusually thin external validation. The engineering is inspectable — the runtime is on GitHub, the weights are downloadable, the benchmarks are documented well enough to reproduce. The claims are not yet reproduced by anyone.

The half of the voice loop Phonon does not cover

A local TTS model is one component of a voice assistant, and it is the last one. Something still has to decide what to say: transcribe the user, retrieve the answer, draft the reply. Phonon turns text into audio on the device, but the text usually comes from a model server, and that server is where the operational complexity actually lives.

That is the boundary OrcaRouter sits on. Phonon is not in our catalogue — it runs on your hardware, not through an API, and nothing here should be read as claiming otherwise. But the language model that writes Phonon's script can be, alongside more than 200 others behind one key, with 0% markup on provider list price and automatic failover when a provider degrades. For offline-first products the split is natural: speech synthesis stays on the device, text generation routes to whichever model is cheapest and healthiest that hour.

What to watch over the next month

Three signals will settle most of the open questions. The first is an independent evaluation — Artificial Analysis adding a speech entry, or a lab rerunning Seed-TTS against the published checkpoints. The second is the commercial licence for the 40M models; if Gradium opens those up, the multilingual checkpoint becomes far more interesting than the English one. The third is voice cloning documentation. Gradium's own blog promises blind listening results "in a future post," which is an honest thing to publish and also an admission that the current evidence is the company grading its own homework.

Until then, the defensible position is this: Phonon 31M is real, it is small, it runs where the vendor says it runs, and it is Apache 2.0. The exact speed and accuracy are somewhere inside the range the vendor's own documents describe. Anyone planning a product around the 7 ms figure should test it on their target hardware first — and anyone planning around the 40M model should read the licence before they fall in love with the voice.