ai-coustics Agent Skills

Skills that teach AI coding agents (Claude Code, Codex, Cursor and others) how to use ai-coustics speech-enhancement and audio-insight models through the AIC SDK.

Skills follow the Agent Skills format: each one is a directory with a SKILL.md that the agent loads on demand, plus the scripts and references it needs.

Install

npx skills add ai-coustics/skills                                  # interactive picker
npx skills add ai-coustics/skills --skill ai-coustics-speech-enhancement   # one skill
npx skills add ai-coustics/skills --skill ai-coustics-audio-insight --agent claude-code

Claude Code users can also install the bundle as a plugin:

/plugin marketplace add ai-coustics/skills
/plugin install ai-coustics-skills@ai-coustics

Manual install: copy skills/<name> into ~/.claude/skills/ (Claude Code) or the equivalent skills directory of your agent.

Prerequisites

Every skill here runs the AIC SDK locally. You need:

• uv on PATH. Each script declares its dependencies inline and runs with uv run.
• An AIC SDK license key exported as AIC_SDK_LICENSE_KEY (the scripts also accept AIC_LICENSE_KEY and AIC_SDK_LICENSE). Generate a self-service key at https://developers.ai-coustics.com.

Models download on first use to ~/.cache/aic-models. Audio never leaves your machine.

Skills

One skill per ai-coustics.com product line.

ai-coustics-speech-enhancement

Enhance speech recordings, single file or whole folder, with Quail Voice Focus, Quail Multi Speaker or Rook Multi Speaker. Async worker pool, latency-compensated output of identical length, mono at the input sample rate.

Use when:

• "Enhance this recording with the AIC SDK"
• "Enhance every call under ./raw/"
• "Apply voice focus at level 0.7"

ai-coustics-audio-insight

Score audio with the Tyto model: a headline risk_score plus per-window quality dimensions such as noise, reverb, interfering speech, codec degradation and packet loss, read from whatever the installed SDK exposes. Predicts whether client-side audio will break VAD, turn-taking, STT or speech-to-speech downstream.

Use when:

• "Score this call with Tyto" / "What's the risk score?"
• "Rank this folder of calls for QA review"
• "Generate the JSON for the call-analysis dashboard"
• "Our LLM judge blames the agent; was it the audio?" (writes judge-prompt context with degraded intervals)
• Explaining what a Tyto dimension means

ai-coustics-voice-activity-detection

Detect speech with the VAD Multi Speaker or VAD Voice Focus models: per-block probabilities, post-processed decisions, and speech segments with delay-compensated timestamps, as a table, CSV or JSON.

Use when:

• "Find the speech segments in this recording"
• "How much of these files is actually speech?"
• "Tune the VAD so our agent stops interrupting callers"
• "Detect only the main speaker, ignore the TV in the background"

Repository structure

skills/                   one directory per skill, kebab-case, ai-coustics- prefixed
  <name>/
    SKILL.md              frontmatter (name, description) + agent instructions
    scripts/              uv-runnable scripts with inline dependencies
      _aic_models.py      shared model-discovery helper (identical copy per skill)
    references/           deep material loaded on demand (optional)
    evals/evals.json      trigger + behaviour test cases
    skill-card.md         owner, credentials, output, risks
plugins/ai-coustics-skills/       Claude Code plugin bundling the catalog (symlinks into skills/)
.claude-plugin/           marketplace manifest
skills.sh.json            grouping for the skills.sh directory
scripts/validate_skills.py  frontmatter, naming, pin and helper-sync checks (runs in CI)
tests/                    pytest for the scripts, heavy deps stubbed in conftest.py

Contributing

See CONTRIBUTING.md for the checklist and AGENTS.md for the authoring conventions agents follow when adding a skill here.

License

MIT for the skills and scripts in this repository. Use of the AIC SDK and model artifacts is governed by the ai-coustics terms.