pocket-tts-onnx-mirror

Pocket TTS ONNX — Post Voice mirror

Version-pinned mirror of the five language bundles used by the Post Voice WordPress plugin, which runs text-to-speech entirely in the reader's browser via ONNX Runtime Web.

Contents

Five languages — english_2026-04, german, italian, portuguese, spanish — each ~190MB:

Only the INT8 quantized models are mirrored — those are the ones the browser runtime loads. The fp32 variants that exist upstream are omitted deliberately: they are never fetched, and carrying them would roughly double the mirror for no benefit.

Why this mirror exists

The plugin pins its model URL to a specific commit of this repository rather than fetching from a third party's repo at main. Pinning a commit protects against file contents changing underneath a released plugin; mirroring protects against the upstream repo becoming unavailable. Bumping the pinned commit is a deliberate, reviewed change, never automatic.

Attribution and license

Everything in this repository is licensed CC-BY-4.0 (Creative Commons Attribution 4.0 International), inherited from both upstream sources:

• kyutai/pocket-tts — the original Pocket TTS model. Declared cc-by-4.0.
• KevinAHM/pocket-tts-onnx — the ONNX export these files come from. Declared cc-by-4.0, and its bundled LICENSE is the Attribution 4.0 text.

The offline Python conversion tooling in the upstream export repo is not mirrored here — it never runs in a browser and is of no use to the plugin.

voices.bin is generated from the upstream bundles by the export_voice_bins.py script that ships with the Pocket TTS web demo; it is not distributed upstream.

A note on a licensing discrepancy found upstream: the Pocket TTS web demo Space distributes a file named ONNX-LICENSE containing the Creative Commons Attribution-NonCommercial 4.0 text. That text conflicts with the Space's own README metadata (license: cc-by-4.0), with kyutai/pocket-tts, and with KevinAHM/pocket-tts-onnx — all three of which declare plain CC-BY-4.0, and the export repo's own LICENSE file contains the Attribution 4.0 text. This mirror therefore carries the Attribution 4.0 license that the actual model sources declare, and does not redistribute the NonCommercial file.

Acceptable use

The CC-BY-4.0 grant on the upstream weights carries a use restriction that follows this mirror: do not use these models for unlawful or deceptive purposes, and do not clone a person's voice without that person's consent.