The Voice Clone Pipeline

• The Reel Factory· The protocol that built the reels you're watching.
• The Sample Lawyer· The clearance pipeline for the sample you already used.
• AI plays Ableton· Claude controls Live through MCP. Sixteen tracks built by talking.

xml
<speak version="1.1" xml:lang="en-MY">
  <prosody rate="98%" pitch="-1st">
    There is a perfectly legal way to outsource this.
    <break time="280ms"/>
    Nobody tells you, because nobody profits from telling you.
  </prosody>
  <break time="600ms"/>
  <prosody rate="103%" pitch="-2st" volume="+1dB">
    Here is how it works.
  </prosody>
</speak>

• SSML prosody markup
• audio mastering (dynamic range, LUFS targeting)
• phoneme-aligned captions (WebVTT / SRT)
• ElevenLabs voice-clone consent ethics
• the uncanny valley in prosody
• right-of-publicity statutes
• the difference between cloning your voice and faking someone else's

If a name is unfamiliar, that's the gap. The list is the curriculum.

1. 01

   Record 90 seconds of your own clean speech. Read varied sentences (statements, questions, lists). Clone once.

2. 02

   Write the script in your written register. The clone preserves your typed cadence better than your spoken one.

3. 03

   Convert prose to SSML. Rate around 95-103%; pitch ±1-2 semitones at section breaks; break tags before commas turn the rhythm human.

4. 04

   Render via ElevenLabs (turbo for drafts; v2 for finals). Generate two takes per line; pick the one that drops the AI tell.

   python
   elevenlabs.text_to_speech.convert(voice_id, text=ssml, model_id='eleven_turbo_v2_5')

5. 05

   Master to -16 LUFS for streaming, -19 LUFS for podcast. Compress lightly; the clone runs hot in the upper mids.

   bash
   ffmpeg -i raw.wav -af 'loudnorm=I=-16:LRA=11:tp=-1.5' mastered.wav

6. 06

   Generate phoneme-aligned captions from the rendered audio (Whisper word_timestamps=true). Trim to <=42 chars/line.

7. 07

   Mux into Remotion or final cut. Run one human listen-through; flag any line that uncanny-valleys and re-render that line only.

Most creators avoid talking-head video because of mic anxiety, schedule friction, or imperfect voice. A trained clone removes all three constraints at the cost of one prerequisite that nothing else does: SSML literacy. The clone is only as alive as the markup; flat SSML produces flat audio. Treat the markup as the script and the script as the performance. Master at the right LUFS and the result reads as podcast-grade.

• The Reel Factory↗
  The protocol that built the reels you're watching.