This workspace has a voice clone prep pipeline designed around one idea:
If you can’t reproduce it, you can’t trust it.
The goal is to take messy, long-form recordings and turn them into clean speaker_01 clips suitable for voice cloning or consistent narration.
Design constraints
- Deterministic filenames and outputs (so runs can be compared)
- Reversible pruning (we flag clips; we don’t destroy them)
- Short segments (e.g. 3–8 seconds) to keep training clean
The pipeline (high level)
- Segment: diarize + export candidate clips
- Auto-prune: score and flag low-quality segments (aggressive mode is reversible)
- Concat: produce a single clean WAV
- (Optional) Stitch: combine multiple runs into a master training file
Canonical commands
See the Atlas module for the exact commands:
atlas/voice-clone-prep.md
Outputs you should expect
A “good” run produces:
segments_all.csv(everything)segments_all.autoprune.csv(keep flags)speaker_01_clean.wav(your training-ready output)
Next: narrated tutorials
The plan is to use the resulting voice clone to generate cached tutorial audio (ElevenLabs), saved under site/public/audio/... and reused across visits.