Local voice cloning director

Local Voice Cloning Director

Turn an authorized reference recording into a reproducible local voiceover. Treat “free” as no per-call API fee; hardware, setup, storage, electricity, and operator time still have costs.

Enforce the rights gate

Read references/safety.md before using any reference recording.

Proceed only when the speaker is the user or the user has documented permission covering voice cloning and the intended use. Refuse public-figure imitation, impersonation, fraud, authentication bypass, harassment, deceptive political persuasion, or any request with unclear consent.

Keep the reference audio, transcript, hashes, generated audio, and manifests local. Never commit them to this Skill repository.

Set up the local engine

Read references/setup.md. Use the official CosyVoice repository and a locally downloaded CosyVoice2 model. Do not download models, install system packages, or change environments without the user’s authorization.

Check the environment:

python3 scripts/check_environment.py \
  --cosyvoice-root "$COSYVOICE_ROOT" \
  --cosyvoice-model "$COSYVOICE_MODEL" \
  --cosyvoice-python "$COSYVOICE_PYTHON"

Prepare the reference

Use a clean 6–15 second clip with one speaker, little noise, and an exact transcript. Create a normalized copy without overwriting the source:

python3 scripts/prepare_reference_audio.py \
  --input /private/input.wav \
  --output /private/reference.wav \
  --authorization own

Do not “improve” the transcript. It must match what is actually spoken.

Prepare spoken text

• Rewrite dense prose into short clauses with natural breathing punctuation.
• Preserve factual meaning.
• Keep English product and person names in their native spelling unless testing proves a correction is needed.
• Use a task-level pronunciation map or a private lexicon; do not put a user’s private lexicon in this repository.
• Read references/quality.md for chunking, loudness, and approval rules.

Inspect the job before synthesis:

python3 scripts/generate_voiceover.py \
  --text-file /private/narration.txt \
  --reference-audio /private/reference.wav \
  --reference-text-file /private/reference.txt \
  --authorization own \
  --consent-confirmed \
  --pronunciation-lexicon /private/pronunciation.json \
  --profile natural \
  --output /private/narration.wav \
  --dry-run

Generate locally

Remove --dry-run after reviewing the chunks. Use natural first. Use steady, expressive, warm, serious, or energetic only when the user asks for a different delivery.

"$COSYVOICE_PYTHON" scripts/generate_voiceover.py \
  --text-file /private/narration.txt \
  --reference-audio /private/reference.wav \
  --reference-text-file /private/reference.txt \
  --authorization own \
  --consent-confirmed \
  --profile natural \
  --speed 1.0 \
  --cosyvoice-root "$COSYVOICE_ROOT" \
  --cosyvoice-model "$COSYVOICE_MODEL" \
  --cosyvoice-python "$COSYVOICE_PYTHON" \
  --output /private/narration.wav

The script writes a local provenance manifest next to the audio. Keep it with the private production job, not in Git.

Transfer delivery style without copying identity

If lawful source audio is available, use scripts/analyze_prosody.py to extract broad pace, pause, pitch-range, and energy-range characteristics. Pass the resulting blueprint to synthesis. Use the authorized target voice as the only identity reference.

Never describe this as cloning the source speaker. Copyright, translation, adaptation, and publication rights remain separate gates.

Validate and listen

Run technical QA:

python3 scripts/validate_voiceover.py \
  --audio /private/narration.wav \
  --text-file /private/narration.txt \
  --manifest /private/narration.wav.voiceover.json \
  --output /private/narration.qa.json

Add --transcribe only when the local Whisper CLI is installed. Technical QA cannot judge warmth, identity match, taste, or whether a delivery is appropriate. Require human listening before replacing a preferred voice profile or using the track in a final published video.

Report honest status

Keep these states separate:

1. environment ready;
2. dry run reviewed;
3. audio generated;
4. technical QA passed;
5. human listening approved;
6. inserted into a video;
7. rendered;
8. published.

Never call an intermediate state “finished.”