Installing a text-to-speech engine is easy. Finding a voice you like is not.
Voice models for Piper TTS are scattered across Hugging Face, Kaggle and various other repositories. Each repository has its own naming conventions, nothing indicates where a model comes from, and licences vary from one resource to another. To choose with confidence, you have to check everything by hand.
PiperHub is Fox3000foxy's project that addresses this problem. A free, open-source catalogue, with no server, that gathers together what was scattered.
What the catalogue lists
The current figures:
- 493 models
- 3,663 speakers
- 70 languages
The catalogue distinguishes the official Piper models from those coming from the community. A useful distinction: the former are published by the maintenance team, the latter come from contributors whose method and data vary.
An audio sample for every entry
Every model has an audio sample. You listen before you download.
This is the most useful feature in practice. A model weighs several dozen megabytes, and downloading it only to discover the voice does not suit you — age, timbre, delivery — wastes time and disk space for nothing.
The catalogue also lets you generate a sample directly in the browser. The synthesis runs on your machine, with nothing to install.
Provenance, which is the real issue
Every model page links back to its original dataset when that is known. The catalogue distinguishes two cases:
- dataset linked: the model already has its dataset referenced.
- to be trained: the dataset is documented, but the corresponding model is still to be produced.
This link between model and dataset is not a decorative detail. A synthetic voice relies on real recordings: it inherits the legal status of the person whose voice was used. Knowing where the recordings come from determines what you are allowed to do with the model.
The catalogue therefore has an entire section devoted to the 130 referenced datasets, each linked to the models that use it.
Filters rather than an endless list
With 493 models, navigation matters as much as content. The catalogue offers filtering by language, by quality (high, medium, low), by number of speakers, and by update date. You can sort by name, by date or by speaker count. An RSS feed announces new voices as they are added.
The catalogue is only part of the project
PiperHub includes a section of guides covering the whole path, from raw data to a usable voice:
- Preparing a dataset in the LJSpeech format expected by Piper: one WAV
file per sentence, plus a
metadata.csvlinking each file to its transcription. - Training a voice, with guidance on resources and quotas.
- Using a voice day to day, from the command line, from Python, or in Home Assistant.
The README also documents how the catalogue is built: the data comes from a
mirrored scraping repository (models_mapping.csv, datasets_mapping.csv),
the voices are mirrored from Hugging Face, and npm run data regenerates the
JSON files the site consumes.
Contributions are made through pull requests, like on any GitHub repository.
How it is built
The site is built with Astro and runs as a static site, with no server. English is the root; the other versions live under their own language code. The interface is translated into sixteen languages, and the repository welcomes contributions from native speakers for strings that have not been translated yet.
Hosting is handled by GitHub Pages, at GitHub Inc.
This choice of a static site matters for a catalogue: there is no database to maintain, no query to handle, no server to secure. The catalogue is regenerated by a script and serves files.
Licences, the project's red line
PiperHub makes a distinction that many catalogues neglect:
- the site code is under the MIT licence;
- the models, samples and datasets remain under their authors' respective licences.
In other words, the catalogue claims nothing over the resources it indexes. It documents; it does not relicense. Every page links to the original resource and to the licence information that belongs to it.
That is what makes the catalogue usable without ambiguity: knowing that a page appears in the index grants no usage rights over the voice it references.
Who it is for
The catalogue is useful as soon as you step outside the trivial case.
If you simply want a French voice for your assistant, a few clicks will do. PiperHub makes real sense when the question becomes serious: someone localising an application, using character voices for a fan project, or building their own voices from recordings whose rights they hold.
The catalogue then serves as an entry point to understand what exists, what is documented, and what can be used.
Sources
- Fox3000foxy — piperhub.org · Source code — Catalogue, voice and dataset indexes, guides.
- Fox3000foxy — Repository README — Site architecture, data pipeline, licences and translations.
- Fox3000foxy — Legal notice — Publisher, hosting and licence handling.