ElevenLabs workflow for voice cloning podcasts

Recommendation (short):
Use ElevenLabs Custom Voice for cloning, record 10–60+ minutes of high-quality, varied audio at 48 kHz/24-bit WAV mono, collect a signed/recorded consent release before recording, clean and split samples, train the clone, then generate speech and run simple podcast post‑processing (EQ, de-essing, limiter, target -16 LUFS). Use chapter markers in your podcast host or ID3 tags for chaptered episodes.

Why this works:
ElevenLabs provides high-quality synthesis and fast iteration. Good source audio + explicit consent protects you legally and gives the model clean signals to learn prosody and timbre. Post-processing keeps the output podcast-ready and consistent with your show’s loudness spec.

Decision criteria (pick based on budget / skill / output needs):
- Budget: If tight, aim for 10–15 minutes of very clean audio; if premium quality and likeness, record 30–60+ minutes and consider an audio engineer. ElevenLabs pricing affects how many test generations you run.
- Skill level: Beginners can follow the checklist below; advanced users should create phonetic dictionaries for unusual words and batch-process with scripts.
- Team size / workflow stage: Single host — simpler consent and one recording session. Multiple hosts/guests — collect per-person consent and separate sessions.

Practical end-to-end checklist
1) Legal consent (before recording)
- Use a written release: identity, permission to synthesize and distribute voice, permitted uses, transfer/licensing terms, duration, revocation clause, compensation if applicable, signature/date.
- Optional stronger proof: record the signer reading the consent aloud (timestamped WAV), save email confirmation and IP metadata.
- Keep one contract per person and store securely.

2) Recording setup & settings
- Environment: treated room, close mic, pop filter, no reflective surfaces or background noise.
- Mic: dynamic (e.g., SM7B) or condenser with cardioid pattern; consistent mic distance 6–12 cm.
- Interface: clean preamp, disable aggressive processing in interface.
- File format and levels: WAV, 48 kHz, 24-bit, mono. Aim for -12 dBFS RMS peaks around -6 dBFS; avoid clipping.
- Delivery material: scripted reads, unscripted conversational monologues, emotional variants, lists (dates, numbers, technical terms). Include multiple prosodies.
- Duration: minimum 3–5 minutes for a basic voice; recommended 10–30 min for high quality; 60+ min for best nuance.

3) Prepare training clips
- Trim leading/trailing silence, remove breaths and clicks, label clips and attach accurate transcripts.
- Normalize to consistent loudness per clip, but avoid heavy compression; preserve natural dynamics.
- Provide a pronunciation guide for names/terms.

4) Clone training (ElevenLabs)
- Create Custom Voice, upload clips + transcripts. Follow the UI to name the voice and add pronunciation overrides.
- Run test prompts, iterate: adjust dataset (remove noisy clips), add more expressive samples, refine pronunciations.

5) Generation and post-processing for episodes
- Generate audio using your clone; export uncompressed WAV.
- Post-process: gentle high-pass (~80 Hz), subtractive EQ (reduce mud), de-esser, mild compression (2:1), limiter to -1 dBTP, loudness to -16 LUFS integrated for spoken-word podcast.
- If you need chaptering: produce a chapter file or embed ID3 chapter frames (many podcast hosts and editors like Audacity, Reaper, or your host's CMS support this). Create a chapter log (timestamp, title, description) and upload to host.

Best-for / Avoid-if
- Best for: recurring host voice continuity, multi-lingual TTS segments, time-savings for editing and ADR.
- Avoid if: you lack clear legal consent; you need an exact emotional performance that only the person can provide live.

Final practical tips
- Keep an iteration log: sample set, date, results.
- Back up original recordings and consent forms.
- Test on a low-stakes episode before using the clone widely.

Mentioned tools: ElevenLabs (for clone creation) and ChatGPT (optional for writing consent templates or test prompts).