voaice

The voice of an AI, written down and checkable. voaice is the voice stack of mindX: speech in, speech out, measured throughout. It gives a machine a voice, says the realm's names the way their owners say them, and proves which voice made a recording.

This repository is voaice's voice library: 70 open-licensed Piper voices in 29 language families (4.62 GB), kept here so a mindX node holds only the voices it speaks with and fetches any other on demand.

Free for anyone to use

The voices in piper/ are here for anyone to use: download one, run it with Piper, build on it. Each voice keeps its own open licence, the licence of the dataset its speakers recorded, and the folder's official licences list every voice under its licence with a link to the legal code. They are Piper's voices, not ours: credit Piper and the dataset named in each voice's MODEL_CARD.

What voaice is

voaice is the voice peer of faicey (the face); with a .persona they compose aivatar, a person's AI model that looks, speaks and thinks. Staged together the suite is ollywoo, the Hollywood of AI avatars, inside the DeltaVerse, where irecto (the director) deploys the finished personas.

Hear it

• rage.pythai.net: every article read aloud from pre-rendered audio. Press LISTEN for the audio deck; OVERLORD of the DeltaVerse opens in the OVERLORD voice, and Three readers, one voice explains the readers.
• playdocs: any document read in any voice of the cast, with a voice studio, a waveform editor and device rendering. A voice your machine can speak as itself is synthesised and encoded on your own processor, inside a budget you set (processor, memory and graphics, 0–100 % to 18 decimals); the host renders the rest.
• deltaverse.pythai.net/voices: the cast · /listen: the reader.
• The render ledger: every rendered article, measured — duration against word count, bytes against duration, said right or said wrong.

Use a voice

from huggingface_hub import hf_hub_download
onnx = hf_hub_download("PYTHAI/voaice", "piper/en_US/ryan/high/en_US-ryan-high.onnx")
cfg  = hf_hub_download("PYTHAI/voaice", "piper/en_US/ryan/high/en_US-ryan-high.onnx.json")
# echo "Hello" | piper --model <onnx> --output_file out.wav

Or with voaice's own fetcher, which refuses any voice not classed open and checks every file's md5: python3 voaice/scripts/piper_voices_hub.py fetch en_US-ryan-high.

Attribution: these voices are Piper's

Every voice here was trained and published by the Piper project and its contributors: rhasspy/piper (Michael Hansen and contributors; Piper is MIT-licensed). They are redistributed unchanged from rhasspy/piper-voices. Each voice keeps the licence of the dataset its speakers recorded, and the original MODEL_CARD (dataset, licence, training notes) sits beside every model exactly as upstream ships it. When you use a voice, credit the dataset named in its card and Piper. See LICENSE.md (licences grouped) and NOTICE.

Only open licences, and why most English voices are missing

Of the 177 voices in the upstream catalogue, a voice is carried here only if both its own licence and, when it is "fine-tuned from" another voice, its parent's are open: CC0, public domain, CC BY, CC BY-SA, Apache-2.0, MIT, Unlicense or GPL/AGPL. The following are left out:

• non-commercial voices (CC BY-NC);
• the lessac / CSTR Blizzard 2013 research-only voices, and every voice fine-tuned from lessac, which is most of the English ones;
• voices whose licence reads only "See URL" or "Unknown".

The full classification of all 177 voices, with the licence text each decision was read from, is piper-voices.json. Copyleft voices (GPL, AGPL, CC BY-SA) keep their share-alike terms in anything built from them. If a classification is wrong, or you hold an excluded voice whose terms allow redistribution, open a discussion.

The voices

English 11 · Dutch 7 · French 5 · Spanish 5 · Ukrainian 5 · German 4 · Kazakh 3 · Basque 2 · Catalan 2 · Farsi 2 · Greek 2 · Italian 2 · Swedish 2 · Telugu 2 · Urdu 2 · Armenian 1 · Bengali 1 · Chinese 1 · Czech 1 · Estonian 1 · Finnish 1 · Latvian 1 · Lithuanian 1 · Marathi 1 · Nepali 1 · Norwegian 1 · Polish 1 · Portuguese 1 · Slovenian 1

piper/<locale>/<voice>/<quality>/ holds <key>.onnx, <key>.onnx.json and MODEL_CARD: the upstream tree minus the language-family directory, every file matching the md5 upstream publishes in voices.json.