Self-hosted Podcasting 2.0 service that turns saved articles into a personal podcast feed.

Paste a URL or upload a document (PDF including scanned, DOCX, Markdown, text, HTML, or an image), wait a few minutes, and get an episode with cloned-voice narration, chapters, artwork, an LLM-written episode summary, and a WebVTT transcript. Subscribe in Pocket Casts, Overcast, or Apple Podcasts like any other show.

Your reading list, as a podcast you own.

• Why
• Screenshots
• Sample
• Quickstart
• Documentation
• What's in the repo
• Development
• Licensing notes
• LLM disclosure
• Credits

I read too much, I like my hands free on the go, and the existing article-to-audio tools either lock the audio in their app, charge per minute, or use voices that sound like an airport PA. I wanted something that:

• I control fully
• produces a real podcast feed any podcatcher can subscribe to
• uses my own voice (or any voice I have rights to)
• keeps my reading list private

That's what this is. No GPU? It runs on CPU too (including Apple Silicon), just slower.

Home: paste a URL or drop files (up to 20 at once), and they join the feed.

Feed: search and page through your episodes, with inline players, transcripts, and per-episode actions.

Settings: browse one category at a time, or search across all settings.

A short clip of cloned-voice narration.

Download the MP4

You need Docker and docker-compose. The app boots unconfigured: set the LLM provider, feed metadata, and admin password, and upload a reference voice, from the Settings UI after it starts.

git clone https://github.com/ttlequals0/Audicle && cd Audicle
cp .env.example .env
# Set BUILD_VERSION in .env to the release tag you want to run.
docker compose up -d

The web UI is at http://localhost:8000/, and it installs as a PWA on a phone home screen. The Feed page shows the exact RSS URL to paste into a podcatcher.

Full setup, including CPU-only deployment and file permissions, is in Installation.

The full index is docs/README.md. The short version:

backend/        FastAPI app, SQLite, the job pipeline
tts-wrapper/    TTS model server (Chatterbox; separate GPU container)
render/         full-article render sidecar and its restricted egress proxy
frontend/       React + Tailwind operator UI
docs/           documentation and screenshots
data/           runtime artifacts (gitignored: SQLite, MP3, JPG, VTT)
docker-compose.yml

Backend:

uv sync
uv run pytest                              # 1000+ tests, a few minutes
uv run uvicorn app.main:create_app --factory --reload --app-dir backend

Frontend:

cd frontend
npm ci
npm run dev

Run npm test and npm run build from frontend/ before submitting UI changes. The build fails on stale design tokens: edit branding/tokens.json, then run node scripts/sync_design_tokens.mjs from the repository root to regenerate the CSS.

Follow the UI design guide and include before and after screenshots for UI changes.

The tts-wrapper/ and render/ packages have their own test suites (uv run pytest in each). Lint with uv run ruff check from the root; one run covers everything. The OpenAPI schema lives at openapi.yaml; regenerate it with uv run python scripts/dump_openapi.py.

CodeQL runs on every PR through GitHub's default-setup code scanning (there is no in-repo workflow file for it).

The application code is MIT. A few things downstream of it have their own terms:

• Chatterbox is the TTS engine. The chatterbox-tts library and its model weights are MIT, so there's no non-commercial restriction on the model itself. Every output carries Resemble's inaudible PerTh watermark for provenance, with no flag to turn it off.
• Wrapper runtime: the wrapper runs Python 3.11 from python:3.11-slim. Its GPU image uses the CUDA 12.6 torch wheels and requires a host driver that exposes CUDA 12.6 or newer. The backend is separate: Python >=3.13, shipped on python:3.14-slim.

The Audicle name and logo are reserved; see branding/README.md.

This project was developed with AI agents as a pair programmer. It was NOT vibe coded. I'm a systems engineer with 15+ years of professional experience; every architecture and design decision here is mine, not the AI's, and every line the LLMs wrote, I reviewed and tested myself.

The paywall bypass strategies are inspired by Ladder; see Paywalled articles.