Recap Studio

Local studio for making recap videos: collect footage, write the script with Claude, generate the voice-over, and render.

Voice engines (all optional, installed from SETUP.bat)

Engine	Runs	Cloning	Notes
Edge-TTS	online, free	no	many voices
Kokoro	offline, any CPU	no	~350 MB
Pocket TTS	offline, any CPU	yes	English, 8 built-in voices
Chatterbox	offline	yes	Nano 110M / Turbo 350M / Original 500M / Multilingual 500M, [laugh] tags
ElevenLabs / MiniMax	online, paid	-	add your key in the app

Install (Windows)

1. Double-click SETUP.bat. It installs the base tools, then asks which voice engines and models to download.
2. Copy keys.env.example to keys.env if you use ElevenLabs or MiniMax (or paste keys in the app's API keys screen).
3. Double-click START_UI.bat.

Run SETUP.bat again at any time to add engines. Chatterbox needs Python 3.11 or 3.10; Nano is installed from the GitHub master branch because the PyPI release does not include it yet.

Character animation

Add transparent PNG cut-outs in the Characters tab (or characters/ in a project). A segment with "character": "name" shows the character in a corner over the clip; a "commentary": true segment with no footage becomes a solo scene (character centered on a background). A talk pose follows the voice-over. In the Premiere XML and VideoPad project the character is its own video track; the mp4 / TikTok builds bake it in. Fields: character, char_pos, char_layout, char_pose, char_anim, char_enter, char_exit, char_bg, char_scale, char_flip, char_outline, char_shadow, char_x, char_y. Needs Pillow (installed by SETUP.bat).

Pocket TTS voice cloning (one-time setup)

Cloning uses Kyutai's gated weights. Built-in Pocket voices work without this.

1. Accept the terms at https://huggingface.co/kyutai/pocket-tts ("Agree and access repository").
2. Create a Read token at https://huggingface.co/settings/tokens.
3. Paste the token in the app under 🔑 API keys → Hugging Face token (or add HF_TOKEN=hf_yourtoken to keys.env).
4. Run SETUP.bat and pick Pocket TTS again (or venvs\pocket\Scripts\python tts_worker.py --download pocket). It prints voice cloning: READY when it worked.

Reference clip format

Use a 16-bit WAV for voice cloning (5-15 s, one speaker). The app's Upload clip button converts mp3/m4a/flac/ogg to WAV automatically (needs ffmpeg), and non-WAV files dropped into voices/ are converted when you generate.

If Pocket TTS reports soundfile is required to read non-WAV or non-16-bit WAV audio files, either install the helper once (new Pocket installs from SETUP.bat already include it):

venvs\pocket\Scripts\python -m pip install soundfile

or convert the clip yourself:

ffmpeg -i myclip.mp3 -ac 1 -ar 24000 -c:a pcm_s16le myclip.wav

Notes

• Chatterbox output carries Resemble AI's Perth watermark.
• Only clone voices you have the right to use. Reference clips go in voices/ (git-ignored).
• Models and engines live in models/ and venvs/ (git-ignored); do not move the folder after running setup.
• Each engine and model has its own license; check them before redistributing audio or models.