phonon-40m

Gradium's Phonon 40M builds as a rhapsode engine, in English, French, German, Spanish and Portuguese. It is the same adapter and runtime as phonon, with the same dials, no cues and no cloning, on the CPU with no torch.

Its weights are CC BY-NC 4.0: non-commercial use only. For commercial use, phonon is Apache-2.0 and speaks English.

Run on real weights in multilingual and de. Both pass rhapsode-conform on their real weights and speak through the OpenAI shim. en, fr, es and pt use the same adapter and file layout, and have only run against a stubbed runtime.

Installing it

The weights are gated, so there are two steps before the first download:

1. Accept the terms on the Hugging Face page of each build you want. Approval is automatic, and each repository asks separately: multilingual, en, fr, de, es, pt.
2. Log in as the user the server runs as, with hf auth login. In the container, write the token to ~/.cache/huggingface/token; docs/operating.md has the command.

Do not put HF_TOKEN in this engine's env block in the config file. Naming an engine there makes it the operator's, and the install is then refused.

pnpm wizard install phonon-40m

The wizard and the web page show both licences and ask. Through the API, the install names the licence it accepts: POST /engines/phonon-40m/install?accept=CC-BY-NC-4.0&pull=de. Without the login the install succeeds and its pull fails, saying which page to visit.

Which build

Six variants, one model each:

Variant	Speaks	Voices	Download
multilingual (default)	All five	103	about 100 MB
en	English	38	about 85 MB
fr	French	16	about 79 MB
de	German	14	about 78 MB
es	Spanish	26	about 82 MB
pt	Portuguese	13	about 78 MB

A build for one language reads it better. On Gradium's own listening tests de scored 2097 Elo in German and multilingual 2046.

multilingual reads language from the request and keeps one model in memory for all five, with nothing reloaded between them.

language decides how numbers and symbols are spelled out, and nothing else. The model has no language input and reads the words it is given: the same plain words asked for in de and in en were identical audio, sample for sample. So the words have to be in the language you want, and language has to match them, because 5 € in a German sentence read as English says "5 euros" and @ says "at". Nothing reports that mistake. With no language it spells as English, the first it lists, whichever voice was asked for. A voice from another language may read any of them, though upstream advises matching them.

A request that names no variant stays on whichever build is loaded (docs/protocol.md § 4), so after a request for de, one asking for French with no variant is refused rather than reloading. A client using more than one build names its variant every time.

Through the OpenAI shim

model takes engine:variant, so a client that can only be given a base URL and a model name can have German:

curl localhost:8080/v1/audio/speech -H 'Content-Type: application/json' \
  -d '{"model":"phonon-40m:de","input":"Guten Abend.","voice":"Nele","response_format":"wav"}' -o speech.wav

The shim cannot pass language, so phonon-40m alone speaks English. Name a build for the language.

What it costs

Measured on an M5 Pro's CPU, November 2026, median of five. de ran an 8 s line at 60x realtime with the first audio out in 14 ms, and multilingual a 6.4 s line at 60x in 14 ms. Each loads in about 0.1 s. On de, a 4,096-character request was 220 s of audio in 3.3 s, with the first of it out in 17 ms and a peak of 412 MB. multilingual peaked at 450 MB, with all 103 voices having spoken. durationDelta moved an 8.16 s line on de between 5.84 s and 9.36 s across its range, and stops doing anything past it, as on phonon.

Upstream's figures for 40M, on a MacBook Air M5's CPU: 20x realtime on one core and 39x on four. A Raspberry Pi 5 manages 2.2x on one core.