What is Gradium Phonon?
Phonon is a family of small on-device text-to-speech models from Gradium, the Paris voice company founded by Kyutai's co-founders, with weights released on 20 to 21 October 2026. The commercially usable one is Phonon 31M English: 31 million parameters excluding the audio tokenizer, a 64.5 MB 8-bit GGUF file, 24 kHz mono output and Apache 2.0 weights, tokenizer and voices. Sister 40M models for English, French, German, Spanish, Portuguese and a five-language multilingual version are under CC BY-NC 4.0, so non-commercial only.
Phonon builds on Kyutai's Pocket TTS and the Mimi codec: a flow-matching transformer predicts audio in 80 ms frames and streams as it goes. About 100 preset voices come in a companion repository; cloning and voice design are paid features of Gradium's hosted service, not the open weights. The 31M model's Hugging Face page records about 165,000 downloads in the past month.
How much does Gradium Phonon cost?
| Plan | Price | What you get |
|---|---|---|
| Phonon 31M English (self-hosted) | Free | Apache 2.0, commercial use allowed |
| Gradium Free | $0 | 45,000 characters a month, up to 5 custom voices |
| Gradium S | $39/month | 900,000 characters, 1,000 custom voices |
| Gradium L | $1,480/month | 45 million characters, 20 pro voice clones |
Phonon 31M English is free under Apache 2.0; the 40M models are CC BY-NC 4.0. Gradium's hosted API has a $0 tier with 45,000 characters a month, S at $39 a month and L at $1,480 a month for 45 million characters, on annual billing.
Is Gradium Phonon worth it?
Phonon 31M is the strongest free English voice that fits inside an app. On Artificial Analysis's arena it scores 1,114 Elo, 41st of 104 and above Kokoro 82M at 1,064, Chatterbox HD at 1,111 and OpenAI's TTS-1 HD at 1,107, at a listed price of zero. It does this at 31 million parameters, under half of Kokoro's 82 million.
Gradium's own blind test, with 468 raters and 8,290 votes, agrees: 2,098 Elo for the 31M model against 2,043 for Kokoro and 2,034 for NeuTTS. Its error rate on hard sentences full of numbers, dates and codes is 1.5%, against 3.5% for Kokoro and 8.7% for NeuTTS, helped by a runtime that spells out numbers and symbols. Gradium reports 43 ms to first audio on an AMD EPYC 7413 CPU and 5 ms on an NVIDIA L40S GPU.
It reaches more platforms at launch than most open speech models: a WASM build for browsers, Core ML for iPhone and Mac, an Android system voice that works with TalkBack, a Docker server with an OpenAI-style API, Home Assistant, Rust and C. That makes it a practical accessibility component, not only a demo.
The catch: the open model is fixed-voice English. Cloning and voice design sit behind Gradium's paid service, the better-scoring 40M English model and every non-English model are non-commercial, and long inputs are split into sentences, so prosody does not carry across a paragraph. The model card asks users not to impersonate people or clone voices without consent; there is no watermark.
Our take: for any English app that needs to talk offline, Phonon 31M now beats Kokoro on the numbers that matter, and the move costs a 64.5 MB download.
Where it earns its place
An Android developer building an offline reading app
Read e-books aloud without a network or per-character fees
Install Phonon as the system TTS engine or embed the runtime; the 64.5 MB model streams from the phone's CPU. Apache 2.0 allows it in a paid app.
A bank's digital team
Read account numbers, dates and amounts aloud correctly in a browser widget
Load the npm phonon-tts WASM build; the runtime's normaliser expands numbers and dates, and the 1.5% hard-sentence error rate is the lowest in Gradium's comparison.
A smart-home hobbyist running Home Assistant
Replace a cloud voice for announcements
Run Gradium's Docker server with the Home Assistant setup on a small CPU box; 43 ms to first audio on a server CPU means announcements start almost at once.
A start-up that needs a cloned brand voice in English and French
Get the same voice across both languages
The open weights cannot clone, so this means Gradium's hosted API: the S plan at $39 a month includes 900,000 characters and instant clones from about 10 seconds of audio.
Before you commit
- Setup
- pip install ptts or run the ptts CLI via uvx or Homebrew; weights and voices download from Hugging Face. Set lang to match the text.
- Integrations
- Python, Rust, C and C++, Swift with Core ML, npm phonon-tts for browsers, Android system TTS engine, Docker server with OpenAI-compatible API, Home Assistant, WebSocket streaming, Transformers.
- Data and privacy
- On-device inference keeps text and audio on the device. The open model has no cloning; the card asks users not to impersonate people or clone voices without consent. No watermark.
- Cost at scale
- Free for the 31M English model. Gradium's hosted API, needed for cloning or other languages commercially, costs $39 a month for 900,000 characters up to $1,480 for 45 million.
- Limits
- English only for the commercial model; apart from the five-language 40M model, each model speaks one language; long text is split by sentence; the Transformers tokenizer does not expand numbers the way the runtime does.
What’s good
- 1,114 Elo on Artificial Analysis, above Kokoro at 1,064
- 31M parameters, 64.5 MB at 8-bit, runs on CPU and in browsers
- 1.5% error rate on hard sentences with numbers and dates
- Apache 2.0 weights and voices; Android, iOS, web and server runtimes
What’s not
- Open model is English only with fixed voices
- Cloning and voice design only through Gradium's paid API
- 40M and non-English models are non-commercial
- No watermark; prosody resets at each sentence
Best for
App and device makers who need a natural English voice that runs offline on phones, laptops or in a browser, under a commercial licence.
Gradium Phonon alternatives
Gradium Phonon: common questions
- Is Gradium Phonon worth it?
- Phonon 31M is the tiny English voice to ship in an app today: smaller than Kokoro, preferred by listeners, and Apache-licensed. We rate it 4.1 out of 5.
- How much does Gradium Phonon cost?
- Phonon 31M English is free under Apache 2.0; the 40M models are CC BY-NC 4.0. Gradium's hosted API has a $0 tier with 45,000 characters a month, S at $39 a month and L at $1,480 a month for 45 million characters, on annual billing.
- What are the downsides of Gradium Phonon?
- Open model is English only with fixed voices; cloning and voice design only through Gradium's paid API; 40M and non-English models are non-commercial; no watermark; prosody resets at each sentence.
- Who is Gradium Phonon best for?
- App and device makers who need a natural English voice that runs offline on phones, laptops or in a browser, under a commercial licence.
- How much does Gradium Phonon cost for a team or in production?
- Free for the 31M English model. Gradium's hosted API, needed for cloning or other languages commercially, costs $39 a month for 900,000 characters up to $1,480 for 45 million.
- Is Gradium Phonon safe for company data?
- On-device inference keeps text and audio on the device. The open model has no cloning; the card asks users not to impersonate people or clone voices without consent. No watermark.
- What are the best Gradium Phonon alternatives?
- The closest alternatives we have reviewed are Kokoro-82M, NeuTTS (Air, Nano, 2E), KittenTTS 2 and Chatterbox.
Answers checked Nov 7, 2026.
Make Gradium Phonon? Show this rating on your site
