piper-voices-ansulev

A pinned mirror of four voices from rhasspy/piper-voices. Not a fork, not a re-training, not an improvement — the same bytes, kept reachable.

Why this exists

rhasspy/piper-voices is a git-lfs repository with no release tags. Anything that pins its files by hash — an Arch PKGBUILD, a Nix derivation, a lockfile — depends on the LFS object at a given path never changing. If upstream re-pushes an object under the same path, every pinned hash breaks at once and there is no way to recover the bytes the pin was made against.

That is the failure this mirror addresses. Upstream vanishing entirely is the less likely risk; a silent re-push is the realistic one.

Upstream remains the source of truth. New voices, new quality tiers and fixes appear there and will never appear here on their own. This repository is an archive of a specific selection at a specific moment, not a competing catalogue. If you are choosing a voice, choose it upstream.

Contents

Directory layout matches upstream exactly (<lang>/<locale>/<voice>/<quality>/), so URLs are interchangeable by swapping the repo name.

Voice	Locale	Quality	Speakers	Note
sharvard	es_ES	medium	2 — "M"=0, "F"=1	es_ES has no high tier upstream; medium is the ceiling for peninsular Spanish
upc_ona	ca_ES	medium	1 (female)	upc_pau (male) exists only at x_low
lessac	en_US	high	1 (female)	professional studio corpus (Blizzard 2011)
ryan	en_US	high	1 (male)	the only en_US male voice with a high tier

All four are 22050 Hz — medium and high alike. Only the low/x_low tiers are 16000 Hz, and none are mirrored here.

sharvard holding two speakers in one file is easy to miss: a consumer that passes only --model will always get speaker 0 (male) and the female voice is unreachable. Pass --speaker 1.

Integrity

Verify against upstream, not against this file — the point of a mirror is that it can be checked.

fdb652db8c11a4475527346cf3241cb064d1ba393cf370f3f2ec09a872d118fd  ca/ca_ES/upc_ona/medium/ca_ES-upc_ona-medium.onnx
7f76acc9c06f4eda9e6aef2997b75782d97855aab48d4b401eb956a6e655eddc  ca/ca_ES/upc_ona/medium/ca_ES-upc_ona-medium.onnx.json
4cabf7c3a638017137f34a1516522032d4fe3f38228a843cc9b764ddcbcd9e09  en/en_US/lessac/high/en_US-lessac-high.onnx
db42b97d9859f257bc1561b8ed980e7fb2398402050a74ddd6cbec931a92412f  en/en_US/lessac/high/en_US-lessac-high.onnx.json
b3990d7606e183ec8dbfba70a4607074f162de1a0c412e0180d1ff60bb154eca  en/en_US/ryan/high/en_US-ryan-high.onnx
c6d3b98f08315cb4bebf0d49d50fc4ff491b503c64b940cd3d5ca28543b48011  en/en_US/ryan/high/en_US-ryan-high.onnx.json
40febfb1679c69a4505ff311dc136e121e3419a13a290ef264fdf43ddedd0fb1  es/es_ES/sharvard/medium/es_ES-sharvard-medium.onnx
7438c9b699c72b0c3388dae1b68d3f364dc66a2150fe554a1c11f03372957b2c  es/es_ES/sharvard/medium/es_ES-sharvard-medium.onnx.json

Usage

model=en/en_US/lessac/high/en_US-lessac-high.onnx
echo "Hello world" | piper-tts --model "$model" --output-raw \
  | paplay --raw --rate=22050 --format=s16le --channels=1

# Spanish female: speaker 1 of the two-speaker sharvard model
echo "Hola" | piper-tts --model es/es_ES/sharvard/medium/es_ES-sharvard-medium.onnx \
  --speaker 1 --output-raw | paplay --raw --rate=22050 --format=s16le --channels=1

Licence and credit

MIT, unchanged from upstream. Every model here was trained and published by the piper project and its voice contributors; the upc_ona Catalan voice originates from Universitat Politècnica de Catalunya. No claim is made over any of it.

Mirrored 2026-09-11.