Last updated: 9 November 2026
Metaphona is built on open models, voices and fonts. Their authors publish them under licences that allow this use, and several require credit: it is given here, as on the About page.
Open models and voices
Each model and voice below is used as its authors publish it, under its licence. Licences of the Creative Commons Attribution (CC BY) family require the author and the licence to be named, with a link to the source.
Used for: Spanish transcription
Whisper large-v3-turbo (faster-whisper)
OpenAI; SYSTRAN
License: MITgithub.com(opens in a new tab)Used for: Speech / music separation
Demucs (htdemucs)
Alexandre Défossez, Meta AI
License: MITgithub.com(opens in a new tab)Used for: Speaker recognition
SpeechBrain ECAPA-TDNN (VoxCeleb)
SpeechBrain project
License: Apache 2.0huggingface.co(opens in a new tab)Used for: Spanish → French translation
OPUS-MT es-fr
Jörg Tiedemann et al., University of Helsinki
License: CC BY 4.0huggingface.co(opens in a new tab)Used for: English → Portuguese translation
OPUS-MT tc-big-en-pt
Jörg Tiedemann et al., University of Helsinki
License: CC BY 4.0huggingface.co(opens in a new tab)Used for: Spanish → Portuguese translation
OPUS-MT tc-big-itc-itc
Jörg Tiedemann et al., University of Helsinki
License: CC BY 4.0huggingface.co(opens in a new tab)Used for: French voice (default)
Piper voice fr_FR-siwis-medium
SIWIS French Speech Synthesis Database; Rhasspy
License: CC BY 4.0huggingface.co(opens in a new tab)Used for: French voices (four speakers)
Piper voice fr_FR-mls-medium (speakers 6856, 10065, 12899, 6128)
Multilingual LibriSpeech (Pratap et al.); Rhasspy
License: CC BY 4.0www.openslr.org(opens in a new tab)Used for: French voice (alternative)
Piper voice fr_FR-gilles-low
French single-speaker speech dataset (Kaggle); Rhasspy
License: CC0 1.0huggingface.co(opens in a new tab)Used for: French voice (alternative)
Kokoro-82M, voice ff_siwis
hexgrad; SIWIS French Speech Synthesis Database
License: Apache 2.0 (voice data CC BY 4.0)huggingface.co(opens in a new tab)Used for: Cloned voices (main model)
Fun-CosyVoice3-0.5B (CosyVoice 3)
FunAudioLLM, Tongyi Lab (Alibaba)
License: Apache 2.0huggingface.co(opens in a new tab)Used for: Cloned voices (alternative model)
Chatterbox Multilingual
Resemble AI
License: MIThuggingface.co(opens in a new tab)Used for: Speech synthesis engine
Piper
Open Home Foundation (Rhasspy)
License: GPL 3.0 or latergithub.com(opens in a new tab)
The speech synthesis engine
Piper, the engine that speaks the standard voices, is licensed under the GNU GPL, version 3 or later. It runs unchanged on our server and is not part of the web app you download; its source code is available from its authors.
Cloning models
Cloned voices are made by CosyVoice 3 (Apache 2.0) or Chatterbox Multilingual (MIT), on a GPU server in Europe. Chatterbox adds an inaudible watermark to the speech it makes, and we keep it.
AI translation
When AI translation is set up, lines are translated by OpenAI, a paid service used under its own terms; it is not an open model. Only text is sent, never the video or the sound. Otherwise, the built-in Opus-MT translators listed here run on our server.
Fonts
The interface uses fonts published under the SIL Open Font License 1.1. They are served from our own site through the Fontsource packages.
Used for: Text
Plus Jakarta Sans
The Plus Jakarta Sans Project Authors (Tokotype)
License: SIL OFL 1.1github.com(opens in a new tab)Used for: Headings
Bricolage Grotesque
The Bricolage Grotesque Project Authors (Atelier Triay)
License: SIL OFL 1.1github.com(opens in a new tab)Used for: Arabic, Devanagari, Hebrew, Japanese, Korean and Chinese text
Noto Sans Arabic, Devanagari, Hebrew, JP, KR and SC
The Noto Project Authors
License: SIL OFL 1.1github.com(opens in a new tab)
Keeping this list current
This list follows the About page in the app and changes when a model, a voice or a font is added or removed.