iPod — Podcast → Yoto

A tiny terminal app that turns any kids-podcast RSS feed into a Yoto MYO card — ads trimmed, pixel icons generated, episodes ordered newest-first.

What you get

• 🎧 Pick a feed — curated French starter list (Les Odyssées, Bestioles, Oli, Les P'tits Bateaux, Salut l'info !, Encore une histoire) or your own RSS.
• ✂️ Ad-free audio — auto-detects the first long silence and trims the intro ad.
• 🖼️ Pixel icon per episode — matches against Yoto's native library first, then Openverse (CC images), then Iconify emoji. No hand-picking.
• 🔀 Newest on top — new episodes are prepended to the card; a one-shot reorder action fixes legacy cards.
• ⚡ Quick sync — one keystroke downloads + uploads every episode not already on the card.
• 🟢 Status dots — ● synced, ◌ downloaded, ○ not yet fetched. No re-downloading.

Install

curl -fsSL https://raw.githubusercontent.com/jaudoux/ipod/main/install.sh | bash

Then ipod. On first run, paste your Yoto developer Client ID — the app handles OAuth in your browser.

Needs Python 3.10+, FFmpeg (brew install ffmpeg), and optionally Ollama for smarter icon matching on French titles.

---

Under the hood

Ad trim

pydub.silence scans the first 5 minutes of each episode for the longest silent stretch (≥1s, −40 dBFS floor). Everything before that cut is dropped — which for most French kids podcasts neatly removes the sponsor spot and jingle without touching the story.

Icon matching

For each episode title the matcher walks three sources and stops at the first hit:

1. Yoto native library — asks Ollama for 5 English keyword synonyms, scores every public Yoto icon (tag match = 30 pts, title-word match = 20 pts), boosts the px-tagged native pixel set, penalizes licensed-character icons (BrainBots, Thomas & Friends, Disney, Peppa Pig…). Referenced by mediaId — no upload needed.
2. Openverse — top CC-licensed PNG, center-cropped, Lanczos-downsampled, nearest-neighbored to 16×16, 16-color-quantized, uploaded as a custom icon.
3. Iconify emoji — Noto / Twemoji / Fluent-Emoji / Flat-Color-Icons SVG rendered to 16×16 PNG and uploaded.

Generated icons are cached per podcast in downloads/<podcast>/.icon_cache.json so re-runs are free.

Ollama

When http://localhost:11434 answers, the matcher uses a local chat model (default llama3.1:latest) to turn a French episode title into 5 ranked English synonyms. Non-deterministic, but temperature is kept low (0.3). If Ollama is offline, the matcher falls back to a bundled French→English dictionary plus raw title tokens — matching still works, just less sharply.

Upload pipeline

• Resilient — "downloaded locally" and "synced on Yoto" are tracked separately, so a failed transcode doesn't force a re-download.
• Exponential-backoff polling — after upload, transcode status is polled at 2s, 4s, 8s… capped at 60s over 10 attempts.
• Newest-first — new chapters are prepended, not appended. The 🔀 Reorder action reverses existing cards as a one-shot migration.

Configuration

Environment variables (all optional):

Variable	Default	Purpose
OLLAMA_URL	http://localhost:11434/api/generate	Ollama endpoint.
OLLAMA_MODEL	llama3.1:latest	Model used for keyword extraction.
OLLAMA_TIMEOUT	15	Seconds before the Ollama request gives up.

Local files:

• ~/.config/ipod/podcasts.json — your podcast list. Hand-editable JSON.
• yoto_config.json — your Yoto client ID. (gitignored)
• yoto_tokens.json — OAuth access + refresh tokens. (gitignored)
• yoto_public_icons.json — cached Yoto icon library (~190 KB). (gitignored)
• downloads/ — audio files and per-podcast icon cache. (gitignored)

Project layout

ipod.py          # top-level menu, episode browser, download/upload orchestration
yoto_api.py      # OAuth device flow, content/media/icon endpoints, yoto_menu
icon_factory.py  # keyword extraction, icon matching, backfill
presets.py       # per-user podcast list persisted at ~/.config/ipod/podcasts.json
tui.py           # shared questionary + rich primitives
logo.py          # the ASCII Yoto-Mini banner

Known limitations

• The Iconify fallback needs an SVG→PNG renderer on PATH (cairosvg, rsvg-convert, ImageMagick, or macOS sips).
• Ollama keyword extraction is non-deterministic; unusual French titles can still confuse it — the French→English dictionary and raw tokens are tried as backups.
• A few Yoto icons are licensed content (BrainBots, Thomas & Friends, Disney, Peppa Pig…). The matcher penalizes them heavily but they can still be picked when nothing else matches.

License

MIT — see LICENSE.