Kokoro Agent Voices

Local zero-cost text-to-speech with per-agent voice profiles using Kokoro TTS (82M params). 54 voices available, named agent mappings, WAV output. Use when b...

Security Scan

The skill is internally consistent with its stated purpose (local Kokoro TTS using a Hugging Face model download on first run); nothing requests unrelated credentials or performs unexplained network exfiltration, though there are small implementation notes to be aware of.

Purpose & Capability

Name/description match the included code: the script uses the Kokoro KPipeline and downloads the 'hexgrad/Kokoro-82M' weights from Hugging Face for local TTS. One minor mismatch: the metadata/requirements list espeak-ng and espeak as required binaries, but the included script does not invoke those binaries (it uses Python libraries and macOS 'afplay' only for playback). The espeak requirement appears unnecessary or at least not used by the provided code.

Instruction Scope

SKILL.md explicitly documents that the model is downloaded from Hugging Face on first run (~350MB) and that subsequent inference is local. The script only reads its arguments and writes WAV output (defaults to /tmp). It does not read unrelated files or environment variables, nor does it transmit text/audio to external services beyond the one-time model download.

Install Mechanism

There is no install spec — the skill is instruction- and script-based. That means nothing arbitrary is downloaded/installed by the platform on install time. Runtime dependencies (Python packages like kokoro, soundfile, huggingface_hub, numpy) must be installed by the user/environment; model weights are fetched from a known public Hugging Face repo at runtime.

Credentials

The skill requests no environment variables or credentials. It uses huggingface_hub to snapshot_download a public repo; no auth is required for a public model. There are no unrelated credential requests or config path accesses.

Persistence & Privilege

The skill is not marked always:true, does not request persistent system-wide changes, and does not modify other skills. Autonomous invocation is allowed but is the platform default and not a special privilege here.

Assessment

This skill appears to do what it says: it downloads Kokoro model weights from Hugging Face once (≈350MB) and runs inference locally. Before installing, consider: (1) ensure you have sufficient disk space and are okay with a one-time outbound download from the public 'hexgrad/Kokoro-82M' repo; (2) install the required Python packages (kokoro, soundfile, numpy, huggingface_hub) in a virtualenv to avoid system contamination; (3) the script's shebang references a developer's local virtualenv path — invoke it with your python (python3 scripts/speak.py) rather than relying on the shebang; (4) the declared requirement for espeak/espeak-ng in metadata is not used in the script and can likely be ignored unless you plan to use those tools separately; (5) if you need to ensure absolute offline behavior, pre-download the model into the expected cache (or vendor the model) so the skill will not attempt network access at runtime. If you need more assurance, inspect or run the script in an isolated environment (VM/container) and verify the model source and license on Hugging Face before use.

Give each AI agent a distinct voice using Kokoro TTS — a lightweight 82M parameter model that runs locally with zero API costs. 54 voices across American, British, and other accents.

Agent Voice Profiles

AGENT_VOICES = {
    "loki": "am_fenrir",      # Deep, authoritative
    "archie": "bm_george",    # British analytical
    "sara": "af_bella",       # Warm, creative
    "kit": "am_echo",         # Clear, technical
    "liv": "af_nova",         # Bright, energetic
    "belle": "bf_emma",       # Refined, thoughtful
}

Usage

python3 scripts/speak.py --agent loki "System check complete"
python3 scripts/speak.py --voice af_bella "Hello world" --output /tmp/greeting.wav
python3 scripts/speak.py --list-voices    # Show all 54 voices
python3 scripts/speak.py --list-agents    # Show configured agent profiles

Setup

Requires a Python environment with kokoro, soundfile, and espeak-ng installed. The model downloads automatically from Hugging Face on first use (~350MB).

Files

• scripts/speak.py — TTS script with agent profiles and voice selection