Use Kokoro-82M for local text-to-speech. Prefer the bundled CLI so generation is reproducible and does not depend on MCP.
Quick Start
Run the script with Python:
python "C:\Users\devan\.codex\skills\speech-kokoro\scripts\kokoro_tts.py" generate --text "Hello from Kokoro." --out "output\speech\hello.wav"
Default behavior:
- Voice:
af_sky - Speed:
1.25 - Sample rate:
24000 - Format: mono 16-bit PCM WAV
- Device: CUDA when PyTorch reports CUDA availability, otherwise CPU
Commands
List voices:
python "C:\Users\devan\.codex\skills\speech-kokoro\scripts\kokoro_tts.py" voices
Generate speech:
python "C:\Users\devan\.codex\skills\speech-kokoro\scripts\kokoro_tts.py" generate --text "Welcome to the demo." --voice af_sky --speed 1.1 --out "output\speech\demo.wav"
Generate from a UTF-8 text file:
python "C:\Users\devan\.codex\skills\speech-kokoro\scripts\kokoro_tts.py" generate --text-file "script.txt" --out "output\speech\narration.wav"
Workflow
- Preserve the user's text exactly unless they ask for edits.
- Use
af_skyunless the user names another Kokoro voice. - Use
voiceswhen the user asks what voices are available. - Write final files to a clear project output path, usually
output\speech\. - For long text, let the script segment and concatenate the audio.
- Report the output path, voice, speed, and whether CUDA or CPU was used.
Dependencies
Install only when missing:
python -m pip install kokoro numpy soundfile huggingface_hub
PyTorch is optional but recommended. The script will use CPU if CUDA is unavailable.
On Windows, Python 3.13 may force native spaCy/blis builds while installing Kokoro dependencies. If that happens, create a Python 3.11 or 3.12 virtual environment for Kokoro and run the same script with that interpreter.
Notes
kokoromust be importable for generation.- Voice discovery uses Hugging Face repo metadata when available and falls back to the known Kokoro voice list.
- Do not store user secrets in this skill. Kokoro generation does not require an HF token for normal local use.