kokoro-local-ttslisted

# Kokoro-82M Local TTS (deployment reference) ## Overview Kokoro-82M (hexgrad/Kokoro-82M on Hugging Face) is an 82M-parameter, Apache-2.0 TTS model whose quality approaches commercial APIs at zero marginal cost. It runs comfortably on any Apple-Silicon Mac or modest GPU. This skill is the deployment recipe plus the operational facts a model card never tells you. ## Setup (once per machine) ```bash brew install espeak-ng ffmpeg python@3.12 atomicparsley mp4v2 # Linux: apt install espeak-ng ffmpeg python3.12 mkdir -p ~/tts && cd ~/tts python3.12 -m venv .venv .venv/bin/pip install "kokoro>=0.9.4" soundfile numpy ``` First synthesis auto-downloads the model weights (~330 MB) from Hugging Face into `~/.cache/huggingface`, plus spaCy's `en_core_web_sm`. No account or token needed. Why each piece (the parts the model card assumes you know): - **Python 3.12, not the newest**: torch and friends ship prebuilt wheels one or two versions behind the latest Python. Chasing 3.14 fails at install time. - **espeak-ng even though the model is neural**: misaki (the grapheme-to-phoneme layer, i.e. text → pronunciation symbols) falls back to it for out-of-vocabulary words. Missing it fails at runtime on rare words, not at import. - **ffmpeg**: not for the model itself; for encoding waveforms into AAC/m4b and anything container-related. - The model card's install commands use `!pip` / `!apt-get` prefixes, meaning they were written for Colab (Linux notebook). Translate `apt-get` →