comparativa-tts-catala-base

Base Docker image for the Síntesi en català HuggingFace space.

Why this exists

The HuggingFace space compares several Catalan TTS engines and depends on a set of heavy, slow-to-build system dependencies:

• festival + festvox-ca-ona-hts (Catalan female voice, from Debian repos)
• festvox-ca-pau-hts (Catalan male voice) — this package was only ever available in the ppa:zeehio/festcat Ubuntu PPA, which is unreachable from HuggingFace's build environment. The voice files are installed directly from FestCat/festcat-voices instead.
• espeak-ng compiled from source from the ca-pr branch of projecte-aina's fork, which adds Catalan support not yet in upstream.
• VITS with its compiled monotonic_align C extension.

Building all of this from scratch on every HuggingFace deploy took 10–15 minutes and was fragile: any branch deletion, PPA outage, or base image tag move would silently break the space (as happened in May 2026 when python:3.9 moved from Debian Bullseye to Trixie, removing apt-key).

What this repo does

It builds and publishes ghcr.io/softcatala/comparativa-tts-catala-base via GitHub Actions whenever Dockerfile.base changes. All upstream dependencies are pinned by commit SHA so the image is reproducible and immune to upstream changes.

The HuggingFace space Dockerfile simply does:

FROM ghcr.io/softcatala/comparativa-tts-catala-base:latest

…and then copies the application code on top. Space builds take seconds instead of minutes, and the compiled artefacts are preserved in the base image indefinitely.

Updating a dependency

1. Update the relevant commit SHA in Dockerfile.base
2. Push — GitHub Actions rebuilds and pushes a new latest tag to GHCR
3. Trigger a rebuild of the HuggingFace space