The video is strongest as a transparent build story: a real adapter, a corrected misunderstanding, an unfinished evaluation, and a clearer experiment at the end.
A YouTube and X narrative built around process rather than a victory lap.
01 / Story
The experiment, not the victory lap
The viewer should leave understanding what was trained, what evidence exists, why the renderer changed, and what would count as a real win. The tension is not whether the code ran. It is whether a small supervised dataset changed useful behavior.
I trained Inkling on the delivery of 42 real UGC ads. The adapter exists and the pipeline works. But I do not yet have enough evidence to tell you the voice direction is better.
Claim-safe cold open
Act I
Hear the missing layer
Show how identical words can become different ads through pace, emphasis, breath, and emotional posture.
Act II
Build the training pipe
Select performances, label delivery, quarantine bad rows, and train the adapter.
Act III
Refuse easy certainty
Explain the weak evaluation, the renderer decision, and the better next test.
02 / YouTube
A 10–12 minute narrative spine
Cold open
“I trained an audio-language model to turn how winning UGC ads sound into performance direction. The checkpoint is real. The result is not proven.” Show the four-stage pipeline and the pending evaluation badge.
The missing data
Demonstrate how a flat transcript loses the smile, rush, pause, hesitation, or certainty that made the delivery persuasive.
What Inkling does
Clarify that Inkling reads audio and text, outputs text, and does not synthesize the final voice. Introduce Fish as the downstream renderer.
Select the cohort
Show a sanitized version of the local selection interface: 3,000+ source ads, 50 chosen, no private media or client identity on screen.
Create expressive targets
Use a synthetic example to show source words, global direction, segment cues, and the word-preservation gate.
Train carefully
Explain the one-row proof, then the 42-row cohort: 31 text-only, 11 audio-conditioned, two rank-32 passes, 42 optimizer steps each.
The failed attempt
Show the incident honestly: connection failure before step zero, dependency correction, then the same bounded run succeeding.
What the adapter learned
Explain existing script → same script with expressive direction. Separate that from the broader “write a new ad” prompt.
Why ElevenLabs became Fish
Describe the closed-list project policy, the discovery that it was too restrictive, and Fish's natural-language expressive cues.
Compare without cheating
Show Generate Ads as exploratory and Compare Markup as controlled. Then show the same-voice blind A/B plan—without inventing a winner.
The better next run
End with 100–300 reviewed pairs, 20–30 untouched scripts, separate writing and direction stages, and a declared listening gate.
03 / X cut
A 75–90 second version
00–08s
Hook
“I had 3,000 high-performing UGC ads—and I wanted to train the part normal transcripts erase.” Visual: waveform becomes expressive text.
08–22s
Method
“I selected 50 performances, kept 42 valid labels, and trained a rank-32 Inkling adapter in two passes.” Visual: 50 → 42 → 84 steps.
22–38s
Boundary
“Inkling does not make the voice. It writes performance direction. Fish will render both models with the same voice.” Visual: Inkling text → Fish audio.
38–56s
Twist
“The adapter exists, but I trained mostly script-to-markup—not product-brief-to-great-ad. So a flashy Red Bull prompt is not the clean test.”
56–74s
Evaluation
“The real test locks one script, preserves every word, renders both outputs blind, and reveals identity only after I choose.”
74–90s
Close
“Pipeline proven. Improvement pending. That is the result—and the next experiment is better because of it.”
04 / Safety
Screen-recording safety and publishing checklist
Use the public site, synthetic examples, green test summaries, abstract model diagrams, and redacted dashboards. Do not improvise a terminal tour while authenticated provider state is visible.
Safe to show
Keep off-screen
Aggregate counts, task structure, sanitized UI, test results
Environment files, API values, request headers
Generic base → adapter → renderer diagrams
Private checkpoint, session, or sampler identifiers
Synthetic scripts and expressive examples
Raw ads, client names, source URLs, private transcripts
“Checkpoint saved” and step totals
Local machine paths, hashes, provider logs
Pending comparison interface
A fabricated or prematurely revealed A/B winner
Title and thumbnail directions
Title: “I Trained an Audio Model on 42 UGC Ads—Did It Learn the Delivery?”
Alternative: “Can AI Hear What Makes a UGC Ad Work?”
Thumbnail: one waveform on the left, one expressive script on the right, and the restrained line “PIPELINE ≠ PROOF.” Avoid model logos and exaggerated result claims.
Publishable disclosure
This is an independent experiment using a private ad library and hosted training services. Source media, credentials, and checkpoint identities are not public. Training completion is verified; expressive-quality improvement remains pending a blind listening evaluation.