KIF Pokédex Voice Over

A mod for Kuray's Infinite Fusion (KIF) that reads every Pokédex entry aloud in the style of Dexter, the classic robotic narrator from the original Pokémon anime — for every base Pokémon, every custom dex entry, and every auto-generated fusion (head × body) the game can produce.

The mod ships an offline voice-clone sidecar built on the open-source fish-speech engine — the same engine that powers the Fish Audio Pokédex Voice Over model. After a one-time first-launch install everything runs fully offline, with no API keys and no network calls at play time.

---

Install (plug-and-play, zero prerequisites)

1. Download the mod ZIP from the Releases page (or via KIF's in-game Mod Browser once it's listed at KIF-Mods/mods).

2. Extract so the pokedex_voice_over/ folder lives in your KIF game's Mods/ directory:

   <KIF game root>/
   └── Mods/
       └── pokedex_voice_over/
           ├── mod.json
           ├── main.rb
           └── fish_speech/   ← the offline TTS sidecar
   

3. Launch KIF and open any Pokédex entry. That's it.

That's the entire procedure — there is nothing else to install by hand. The first time you open the Pokédex, the mod auto-spawns the Fish-Speech sidecar in a minimised terminal window, which performs a complete one-time first-run setup in this order:

• Python — Windows uses a bundled portable Python 3.12 that lives entirely inside fish_speech/python/. The launcher extracts it from the official python.org embeddable ZIP — no installer EXE, no SmartScreen prompt, no admin, no PATH changes, no registry edits. Uninstall is just "delete the mod folder." macOS uses Homebrew; Linux uses your distro's package manager.
• PyTorch — installs the matching CPU or CUDA wheel (auto-detected).
• Engine dependencies — installs the small set of inference-only Python packages the engine needs, listed in fish_speech/requirements-runtime.txt. The Fish-Speech engine itself is bundled with the mod (vendored under fish_speech/vendor/ — see Offline & self-contained design below), so setup never clones or downloads it from GitHub.
• Model weights — downloads the ~1.4 GB voice-clone model from the public fishaudio/fish-speech-1.5 Hugging Face repo. No token, no account, no API key required.

The first entry on first launch may stay silent while this runs; from the second entry onward the voice plays normally. Every subsequent game launch skips straight to the server — usually a 5–15 second warm-up.

If you'd rather pre-warm the install so the first entry isn't silent, run the launcher once before opening the game:

• Windows: double-click Mods/pokedex_voice_over/fish_speech/Start_TTS_Server.bat
• macOS / Linux: ./Mods/pokedex_voice_over/fish_speech/start_tts_server.sh

Leave the terminal open. When the game closes, the mod closes it for you.

On Linux & macOS

Auto-start is fully cross-platform: when you launch KIF the mod spawns start_tts_server.sh in the background on its own (same as Windows, and on by default), so in the common case you don't have to do anything.

There is one caveat for the very first launch. Auto-start runs the script detached with no terminal attached, which means:

• If Python 3.10+ is not already installed, the script tries to install it with your package manager via sudo — which cannot ask for a password from a background process, so the silent auto-start cannot bootstrap Python.
• The one-time ~3 GB download (PyTorch + model) runs silently with no visible progress.

So on Linux/macOS it is best to run the launcher once in a terminal for the first install, then let auto-start take over from then on:

cd "<KIF game root>/Mods/pokedex_voice_over/fish_speech"
bash ./start_tts_server.sh

This creates an isolated ./venv inside the mod folder, installs PyTorch and the engine dependencies into it (you will see progress, plus any sudo prompt to install Python), downloads the model, and runs a quick smoke test. Nothing is installed into your system Python. Once it finishes, every later game launch auto-starts the server silently — exactly like Windows.

Prerequisite: Python 3.10+ with the venv module. On Debian/Ubuntu: sudo apt-get install -y python3 python3-venv (Fedora, Arch and openSUSE ship venv with python3). For a clean reinstall later, delete fish_speech/venv/ and fish_speech/.installed, then relaunch.

---

Features

• 🔊 Reads the full Pokédex description aloud the moment you open any entry.
• 🧬 Works for every fused Pokémon combination — head × body — including auto-generated AI fusion entries.
• 🚀 Zero-prerequisite install. No Python install, no pip commands, no manual model download — the first launch auto-installs everything per-user without an admin prompt.
• 🐟 Fish-Speech voice cloning (offline, primary) — clones the Pokédex narrator reference clip and reads any entry on the fly. No pre-generated audio files needed.
• 🗣️ Piper TTS fallback — if you'd rather not use the heavier Fish-Speech stack, drop a Piper voice into Mods/pokedex_voice_over/piper/ and the mod will use that instead.
• ⚡ Cached output — generated WAVs are cached by content hash in Audio/SE/Pokedex/tts_cache/ so repeat visits play instantly.
• 🎵 Music fades while the Pokédex is open and resumes when you exit.
• 🔄 Scroll-aware playback — voice plays as you scroll through entries with the D-pad / arrow keys, not just when you open one.
• ⚙️ In-game settings (volume, enable/disable, mute music, re-read on page return, prefer Fish-Speech, prefer Piper) via the KIF Mod Manager and Stonewall's Mod Settings.
• 🔐 No personal API keys. The model comes from a public Hugging Face repo; no Fish Audio / FakeYou login is required.

---

In-game settings

Settings are exposed both through KIF's Mod Manager (Title Screen → Mod Manager → Installed Mods → Pokédex Voice Over → Settings) and through Stonewall's Mod Settings if you have it installed (Options → Mod Settings → Interface → Pokédex Voice Over).

---

Requirements

---

How it works

The mod is a single Ruby script (main.rb) that hooks KIF's Pokédex scene to capture the exact text being displayed (including auto-generated fusion entries containing the literal POKENAME placeholder, which it resolves to the actual fused name). It then calls a tiny local HTTP server in fish_speech/ (POST 127.0.0.1:7861/tts) and plays the returned WAV through RPGMaker's audio system.

The server is fish_speech/server.py — a thin wrapper around the fish-speech 1.5 engine, which is vendored into the mod at fish_speech/vendor/ and added to Python's import path at startup. It keeps the model loaded between requests so per-entry latency is sub-second on GPU. The model and its tokenizer are downloaded once from the public fishaudio/fish-speech-1.5 Hugging Face repo — no token, no account, no API key required.

If Fish-Speech is unavailable (e.g. the user declined the Python install prompt or the GPU OOM'd) the mod falls through to its built-in Piper path. Piper is the small offline neural TTS used by Home Assistant; drop the matching piper.exe (or piper binary) and a *.onnx voice model into Mods/pokedex_voice_over/piper/ and the mod picks it up automatically.

---

Offline & self-contained design

The Fish-Speech 1.5 engine source is vendored inside the mod at fish_speech/vendor/ (a trimmed, pinned copy of the upstream code). server.py puts that folder first on Python's import path, so the mod always runs the exact engine it shipped with — it never pip installs fish-speech from GitHub or PyPI. First-run setup therefore only installs (a) the right PyTorch wheel and (b) the inference-only leaf packages in fish_speech/requirements-runtime.txt. After that one-time download, nothing reaches the network at play time.

Maintainer note — keep the engine on 1.5. fish-speech 2.x reorganised the package and removed the firefly VQ-GAN module (fish_speech.models.vqgan.modules.fsq.DownsampleFiniteScalarQuantize) that the 1.5 checkpoint instantiates, so a 2.x engine cannot load these weights. Don't bump the vendored copy to 2.x without also rewriting server.py and swapping the model checkpoint. To refresh the vendored source, copy fish_speech/ and tools/ from the upstream v1.5.0 tag into fish_speech/vendor/, keeping the .project-root marker and the per-directory __init__.py files.

---

Uninstalling

Delete the Mods/pokedex_voice_over/ folder. That's all — the bundled Python interpreter, every pip-installed package, the model weights, and the caches all live inside that folder, so there's no system cleanup left over.

(On macOS / Linux, Python was installed via Homebrew or your distro's package manager and stays installed in case other tools need it; remove it via your usual package-manager uninstall flow if you want it gone.)

---

Troubleshooting

---

Credits & License

• fish-speech by Fish Audio — Apache 2.0 (github.com/fishaudio/fish-speech)
• The bundled reference voice clip was generated from the public Fish Audio Pokédex Voice Over model and is included only as the voice-clone target. You can replace it with any other clip you prefer.
• Piper by Rhasspy — MIT (github.com/rhasspy/piper)
• Pokémon Infinite Fusion by Chardub/Frogman
• Kuray's Infinite Fusion by kurayamiblackheart and contributors
• This mod is MIT-licensed. Pokémon trademarks belong to Nintendo / Game Freak — this is a fan project, not affiliated with or endorsed by them.