voice-dna-extract

name	voice-dna-extract
description	Extract voice DNA from video/audio/YouTube, run identity A/B, and freeze versioned locks (Speaker-1, Speaker-2). Local-first; no personal samples bundled. Part of Voice DNA Studio by Navid Taheri / Navid AI ToolBox.

Voice DNA Extract (public skill)

Use this skill when the user wants to:

• pull a clean clone reference from a video, recording, or YouTube URL
• A/B refs x Chatterbox params for identity stickiness
• freeze a versioned lock (Name-1) without overwriting prior locks

Repo / package

This skill assumes the Voice DNA Studio repo is available:

voice-dna-studio/
  voice_dna_studio/   # python -m voice_dna_studio ...

Set VOICE_DNA_ROOT to the repo path, or cd there before commands.

Hard rules

1. Never commit real DNA audio.
2. Never overwrite voices/<slug>/<Name-N>/ - bump version instead.
3. Prefer soft HP + peak norm; avoid heavy loudnorm as default.
4. Primary DNA window ~10-16 s pure single-speaker speech.
5. No personal absolute paths in docs or defaults.

Workflow

1) Extract

cd "$VOICE_DNA_ROOT"
source .venv/Scripts/activate   # or .venv/bin/activate

python -m voice_dna_studio extract \
  --input "/path/to/source.mp4" \
  --out "workspace/kits/<job_id>"
# or
python -m voice_dna_studio extract \
  --youtube "https://www.youtube.com/watch?v=ID" \
  --out "workspace/kits/<job_id>"

Primary ref: workspace/kits/<job_id>/ref/ref_dna_primary.wav
Meta: DNA_META.json

2) A/B (before long jobs)

python -m voice_dna_studio ab \
  --ref workspace/kits/<job_id>/ref/ref_dna_primary.wav \
  --ref workspace/kits/<job_id>/ref/ref_dna_10s.wav \
  --out workspace/ab/<job_id> \
  --line "Short neutral line in the target language and tone."

Read ranking.csv, listen top takes, pick params.

3) Lock

python -m voice_dna_studio lock \
  --voices-root workspace/voices \
  --slug <speaker_slug> \
  --name <Speaker-1> \
  --ref workspace/kits/<job_id>/ref/ref_dna_primary.wav \
  --exaggeration 0.36 \
  --cfg-weight 0.62 \
  --temperature 0.58 \
  --notes "optional free text, no PII required"

Deliverables to report

• path to primary DNA WAV
• density / start_sec from meta
• top A/B tags + params
• lock folder path + ACTIVE pointer

Anti-patterns

• Shipping generic "female/male" knobs as final without A/B when identity was demanded
• Overwriting locks
• Putting user home directories or private sample names into the public skill text