Phonon 31m is 65 MB, Apache 2.0, and English only
The only Apache 2.0 variant of Phonon speaks one language, and at 65 MB it runs on-device with no server and no network.
| Phonon 31m | Phonon 40m | |
|---|---|---|
| Model size | ~65 MB | ~75 MB |
| License | Apache 2.0 | CC BY-NC |
| Languages | English | English, French, German, Spanish, Portuguese |
| Architecture | flow matching | flow matching |
| Voices | 100+ | 100+ |
The 65 MB figure does not confirm the model fits a phone's memory budget. The document names phones, laptops, and single-board computers but gives no minimum hardware spec for any of them.
What actually gates a commercial product is the license, not the footprint. The 31m variant is the only one embeddable without a per-user fee, and it covers English alone. The 40m variant adds four more languages at roughly 75 MB but carries a CC BY-NC license, so a product needing French or German pays per user for 10 MB of extra capacity.
Flow matching from the Pocket TTS lineage is what makes the 65 MB size plausible. The document reports no inference latency, no per-utterance cost, and no Open TTS Leaderboard score despite claiming to top the benchmarks.
No number backs that claim.
The license split is the real story, not the footprint. The missing benchmark number undercuts the headline claim and should be flagged as unconfirmed.
After arxiv.org. We did not report this. The pictures, if any, are theirs.
