← Back to Voxa

Credits

Voxa is built on open models and open-source software. Voice names are Voxa's own; the voices come from the models below.

Voices

• Kokoro-82M by hexgrad: the Studio voices in all nine languages. Apache-2.0
• The Japanese Studio voices Hana, Sakura, Mei and Haruto were trained on readings from the koniwa project (Gongitsune, Tebukuro wo kaini, Nezumi no yomeiri, Kumo no ito). CC BY
• The French Studio voice Camille was trained on the SIWIS French Speech Synthesis Database. CC BY
• Phonon 31m and the Phonon voice catalogue by Gradium: the on-device English voices. Apache-2.0

Software that runs in your browser

• phonon-tts by Gradium: on-device speech. MIT or Apache-2.0
• pdf.js by Mozilla: reading PDFs for audiobooks. Apache-2.0
• lamejs, a JavaScript port of LAME: making MP3 files. It is served unmodified as its own file, lame.min.js, so it can be replaced; source on GitHub. LGPL-3.0
• supabase-js: signing in. MIT

Software on Voxa's servers

• kokoro and misaki by hexgrad: speech and pronunciation. Apache-2.0
• eSpeak NG through phonemizer: pronunciation fallback. GPL-3.0
• PyTorch, spaCy, jieba, pypinyin, pyopenjtalk, fugashi with UniDic-lite, soundfile and NumPy. Permissive licences (Apache, MIT, BSD)