Gradium Phonon review (2026): honest verdict on the open-source on-device TTS — the 31M Apache model, 40M multilingual models, speed, and licensing.
Gradium Phonon is one of the most capable tiny on-device text-to-speech models you can run in 2026 — sub-10 ms first audio on a laptop CPU, many times real time, and small enough to run in a browser or on a Raspberry Pi. The English 31m model is genuinely free under Apache 2.0 and the clearest reason to try it. Two honest caveats hold the score back: the bigger multilingual 40m models are non-commercial (CC BY-NC) unless you license them, and the speed and accuracy claims are vendor-measured with no independent benchmarks yet. As an open voice model it earns a high score; just pick the variant that matches how you will use the audio.
Gradium Phonon is interesting for the same reason Kokoro was: it makes high-quality narration small, fast, and local. Gradium — a Paris voice-AI company built by former Kyutai researchers — open-sourced Phonon on October 20, 2026, a suite of text-to-speech models all under 50 million parameters, designed to run entirely on a device with no cloud. The smallest, Phonon 31m, is about 64 MB and starts speaking in roughly 7 ms on a laptop CPU. That it does this while sounding natural is the whole pitch.
This review scores Phonon as what it is: an open-weight voice model. It is not a content-creation suite, and I don't grade it as one — no captioning, no image or video generation, no scheduling. Where it competes, against other small open TTS models like Kokoro, it competes at the front of the pack on Gradium's own numbers, and the scores below reflect that while flagging where those numbers are still unverified.
Two things anchor the verdict. First, the engineering is real: a 31M model that runs faster than real time on a phone or a single-board computer, in the browser via WebAssembly, is a genuinely useful building block. Second, the fine print matters more than usual here — only the English 31m model is Apache 2.0, while the 40m per-language and multilingual variants are CC BY-NC, so commercial use of their audio needs a license from Gradium. The right model depends entirely on what you're making and whether you're monetizing it.
Everything below reflects Phonon's state as of 2026-10-21 and is based on Gradium's own release materials and model pages. The latency and word-error-rate figures are vendor-reported; no independent third-party benchmarks existed at the time of writing, so confirm current specs, licenses, and performance on the official pages and test on your own scripts before you commit.
Phonon is a family of on-device text-to-speech models from Gradium, released open-source on October 20, 2026. It replaces an earlier, closed ~100M-parameter private beta of the same name with a smaller, openly available set. There are three variants. Phonon 31m is a 31-million-parameter English model, roughly 64 MB at 8-bit quantization, with 37 built-in voices, released under Apache 2.0. Phonon 40m is a 40-million-parameter, roughly 75 MB model offered one per language across English, French, German, Spanish, and Portuguese. Phonon 40m multilingual is a single 40m model covering all five languages with a large built-in voice set. The 40m variants are licensed CC BY-NC, meaning non-commercial use is free but commercial use requires a license from Gradium. Architecturally the models are compact transformers — 6 layers for the 31m, 9 for the 40m, outputting 24 kHz audio. Gradium reports first audio in about 7 ms on a MacBook Air CPU and many times real-time throughput, with comparable sub-10 ms figures on a Snapdragon phone and faster-than-real-time performance on a Raspberry Pi 5; it also claims lower word-error rates than other small open models on deliberately hard sentences (phone numbers, emails, postal codes). The weights are on Hugging Face and load through Transformers, with code at github.com/gradium-ai/xn-ptts offering Rust, Python (CLI), Swift, and JavaScript/WebAssembly integrations, an OpenAI-compatible server, and day-one Apple Silicon support via MLX-Audio; integrations with LM Studio and Ollama are listed too. Voice cloning and voice design are offered free through Gradium Studio and the API. The models run on CPUs, GPUs, and NPUs, and in the browser. What Phonon does not do is write, see, or publish — it turns text into audio and stops there.
Phonon fits three groups cleanly. First, developers building voice features — especially voice agents — who want low-latency, offline synthesis they can embed on a phone, a laptop, or in a browser without a metered API or a cloud dependency; the sub-10 ms first-audio claim is aimed squarely at them. Second, self-hosters and privacy-sensitive teams who need narration that never leaves the device. Third, budget-conscious creators voicing text at volume — articles, faceless-video scripts, accessibility reads — who can work with the English 31m model and are comfortable assembling the tooling around it. It fits poorly for anyone expecting a finished content tool: Phonon makes audio, not captioned video, carousels, blogs, or published posts, and it has no brand-voice-across-a-week layer, no scheduler, and no app in the usual sense. One extra consideration unique to Phonon: if you need multilingual output in monetized content, the only non-commercial-free path is to license the 40m models from Gradium, which changes the "it's free" calculus.
| Dimension | Score | Why |
|---|---|---|
| Voice quality | 4.1 / 5 | Natural narration for a sub-50M model, and Gradium reports lower word-error rates than rival small models on hard text — strong, though the figures are vendor-measured and unconfirmed by third parties. |
| Speed & latency | 4.8 / 5 | The standout trait: about 7 ms to first audio on a laptop CPU and many times real time, with sub-10 ms on a phone and faster-than-real-time on a Raspberry Pi 5. |
| On-device / offline & browser deployment | 4.7 / 5 | Runs fully offline across CPU, GPU, and NPU targets, and in the browser via WebAssembly and WebGPU — about as portable as on-device TTS gets. |
| License & openness | 3.8 / 5 | Only the English 31m model is Apache 2.0; the 40m and multilingual variants are CC BY-NC, so commercial use of the larger models needs a Gradium license. |
| Voice & language coverage | 3.9 / 5 | English 31m plus five languages on the 40m line, with large built-in voice sets; broader than Kokoro on built-in voices but the full language range sits behind the non-commercial license. |
| Voice cloning & design | 4.0 / 5 | Voice cloning and voice design are offered free via Gradium Studio and the API — a step past Kokoro's preset-only base model. |
| Ease of setup & integrations | 4.0 / 5 | Transformers, MLX-Audio, LM Studio, Ollama, an OpenAI-compatible server, and Rust/Python/Swift/JS bindings make it unusually easy to drop in for an open model. |
| Content-workflow scope | 1.5 / 5 | Voice synthesis only — no written content, images, captions, carousels, scheduling, or publishing. Not what the model is for. |
Phonon's pricing story is mostly a licensing story, and that is the part to read carefully. Local inference is free and unlimited, and voice cloning and design are free through Gradium Studio and the API. For the English Phonon 31m model, which is Apache 2.0, that is the whole picture: run it on your own hardware and use the audio commercially at no cost, the same structural advantage over metered commercial APIs that made Kokoro popular. If your workload is English narration at volume, self-hosting the 31m model takes a recurring voice bill to roughly zero.
The asterisk is the 40m line. Those models — the per-language variants and the multilingual model — are CC BY-NC, so they are free for non-commercial use but require a license from Gradium for commercial use. That means the "Phonon is free" headline is only fully true for the 31m English model; the moment you need French, German, Spanish, or Portuguese audio in monetized content, you are into a sales conversation. This isn't a criticism so much as a planning fact: pick the variant that matches your use, and don't assume the whole suite is free the way Kokoro's is.
The honest read: as an open voice model, the Phonon 31m is priced unbeatably — nothing, under a permissive license — and the engineering effort of standing up and maintaining the pipeline is the real cost, as with any self-hosted model. The multilingual models are a different deal, free to experiment with and paid to monetize. Either way, what the price never includes is the content work around the voice: writing the copy, making the visuals, assembling the video, or publishing anything.
| Use case | Fit | Why |
|---|---|---|
| Low-latency voice agents / interactive apps | Strong | Sub-10 ms first audio on-device, offline, is exactly the profile a real-time voice agent needs, and the latency focus is clearly aimed here. |
| Free, local, private English narration | Strong | The Apache 2.0 31m model runs offline on your own hardware with no cloud dependency and no per-character bill. |
| Embedding a voice in an app or browser | Strong | CPU/GPU/NPU targets plus WebAssembly and an OpenAI-compatible server make integration unusually easy for an open model. |
| Multilingual narration in monetized content | OK | The 40m models cover five languages but are CC BY-NC, so commercial use requires licensing them from Gradium. |
| Cloning or designing a custom voice | OK | Voice cloning and design are free in Gradium Studio, though the resulting commercial rights depend on which base model you use. |
| Generating written posts, scripts, or blogs | Weak | Phonon reads text; it does not write it. There is no copy generation or brand-voice layer. |
| Producing visual content (carousels, quote cards, video) | Weak | The model is audio-only; it generates no images or video assets. |
| Scheduling and publishing across platforms | Weak | No scheduler and no social connections — Phonon publishes nothing. |
To be clear where I stand: I run Kompozy, and Kompozy is not a Phonon competitor. Phonon is weights — an open voice model you deploy on a device. Kompozy is a content operation you run from a dashboard. I include this note because people arrive at Phonon from two different places, and only one of them is a Kompozy question. If you are comparing on-device TTS models — Phonon against Kokoro or a hosted API, on latency, quality, and license — benchmark those directly; Kompozy isn't in that race and I won't pretend it is.
The other arrival is the creator who reached for a free local voice because they were trying to make more content, and then found the audio was never the bottleneck. That is the honest line between the two products. Phonon hands you an audio file; it writes no copy, generates no video or images, keeps no brand voice across a week, and publishes nowhere. Kompozy does that part: copy under a Persona Brief, short-form and avatar video (with its own built-in HeyGen TTS), carousels, quote cards, a blog, and a newsletter, scheduled and published from one queue across nine platforms. The clean pairing many creators land on: Kompozy to generate and ship the content, Phonon's 31m model to voice the written outputs into a free, local audio channel. Two tools, two halves — and because Kompozy's avatar video already carries its own voice, they barely overlap.
For voice, yes — it is one of the strongest tiny on-device TTS models available, with sub-10 ms first audio on a laptop CPU, browser support, and a free Apache 2.0 English model. It is not worth it as a content-creation tool, because it generates no written posts, images, or video and publishes nothing; it is a voice model, not a workflow.
The English Phonon 31m model is free under Apache 2.0, including for commercial use. The 40m per-language and multilingual models are CC BY-NC — free for non-commercial use but requiring a license from Gradium for commercial use. Local inference and voice creation are free and unlimited across the suite.
Both are small open TTS models that run locally on a CPU. Gradium reports faster first-audio latency and lower word-error rates on hard text for Phonon, but those are vendor-measured figures without independent benchmarks. Kokoro is Apache 2.0 across the board; only Phonon's 31m model is, with its multilingual models under a non-commercial license. Test both on your own scripts.
Yes. Phonon is built to run on-device across CPUs, GPUs, and NPUs, including phones and a Raspberry Pi 5, and it runs in the browser via WebAssembly and WebGPU. Gradium reports it runs faster than real time on all of these.
Yes. Gradium offers free voice cloning and voice design through Gradium Studio and its API. The commercial rights to the resulting voice depend on which base model you use — the 31m is Apache 2.0, while the 40m models are non-commercial without a license.
No. Phonon generates audio only — narration, voiceover, streaming speech. It does not write posts, make images or video, caption clips, or schedule and publish to any platform. For that you need a content engine like Kompozy, which many creators pair with a local Phonon voice.
Gradium, a Paris-based voice-AI company founded in 2025 by researchers from the lab Kyutai. Gradium sells real-time voice APIs (speech recognition, synthesis, voice cloning, translation) and open-sourced the Phonon on-device TTS models on October 20, 2026.
Developers building low-latency or offline voice features, self-hosters and privacy-sensitive teams, and budget-conscious creators voicing English text at volume with the free 31m model. It fits poorly for anyone whose real need is producing and distributing finished, on-brand content across platforms.