Short answer / recommendation
Clone a co‑host’s voice only with explicit, documented consent and a signed release. Use high‑quality source recordings (preferably several minutes), produce test renders, and require human signoff of every synthetic ad or edit before publishing. If the co‑host is unwilling, unionized, or you can afford regular pickups, do not clone.

Decision criteria (pick one path)
- Clone if: you need frequent edits/ads, want brand consistency, have the co‑host’s written release, and can store audit logs. Good for small teams that publish many short promos.
- Don’t clone if: the co‑host objects, content is sensitive/controversial, legal or union rules block synthesis, or you lack budget/skill to manage quality and approvals.

Practical checklist (step‑by‑step)
1) Pre‑consent conversation: explain purpose, examples, delivery formats, how long voice will be used, compensation, and revocation policy. Record this conversation (with consent).
2) Written release: use a short, plain‑English agreement + signature (see sample below). Send to your legal counsel for jurisdictional tweaks.
3) Recording session for training samples: quiet treated room, high‑quality mic (dynamic e.g., Shure SM7B or a good large‑diaphragm condenser if room treated), pop filter, audio interface. Record mono WAV, 48 kHz, 24‑bit, consistent mic distance. Capture 5–10 minutes of varied speech (tone, pacing, scripted ads, conversational lines). Minimum: 20–30 seconds may work for basic clones but yields lower fidelity.
4) File prep: trim breaths only when necessary (don’t over‑edit), remove clicks, keep natural prosody. Maintain noise floor below −60 dB. Export as uncompressed WAV, label with metadata (speaker name, date).
5) Upload to ElevenLabs and create the voice model. Choose the high‑quality model; if offered, use “similarity boost” or fine‑tune options. For ad outputs use low creativity (temperature 0.2–0.4) and moderate stability so intonation stays natural.
6) Test renders: generate 3–5 short scripts of varying register. Provide the co‑host with side‑by‑side originals and synthetic versions for blind evaluation.
7) Approval workflow: require signed approval for (a) initial model, (b) each ad/campaign, and (c) any change in use. Keep versioned logs of all synthetic files and approvals.
8) Post‑use audit: embed metadata (project ID, date, voice model ID) in delivered files and keep retention records for at least 2 years (or as required).

Sample legal release (plain English, ~2–3 sentences)
“I, [Name], hereby grant [Producer] the non‑exclusive, worldwide right to create and use a synthetic/replicated version of my voice for production, advertising, and editorial purposes. I confirm I consent to the reproduction and distribution of audio containing my likeness as described above, and I understand I will be compensated as agreed. This authorization is effective from [date] until revoked in writing; revocation will not invalidate uses already produced in good faith. (This form does not replace legal advice.)”

Best‑for / Avoid‑if
- Best for: recurring ads, rapid edits, long‑running shows needing consistent reads, small teams needing fast turnaround.
- Avoid if: the co‑host is uncomfortable, legal/union rules restrict synthetic voice, or you require perfect emotional nuance for high‑stakes content.

Notes on budget & team size
Higher fidelity (more training minutes, pro engineer, legal review) costs more. Small teams can use a simple consent + tests; larger productions should add contract attorneys, versioning systems, and dedicated QA. For wording help or iterative drafts, using ChatGPT can speed drafting before counsel reviews.

Final tip
Always keep the human in the loop: signoffs, audible labels (“synthetic voice”), and an easy revocation process maintain trust with your co‑host and your audience.

Compare options for ElevenLabs