Timing
Pauses, rushes, rests, and segment boundaries can be grounded in the recording.
What we were actually trying to teach
The narrow task was expressive direction—not voice cloning, not audio synthesis, and not a magical shortcut to a full creative director.
Inkling, LoRA, cross-modal supervision, and the boundary between writing and direction.
01 / Objective
Teach a transformation we can inspect
The useful target is a performance-ready version of an existing script. The output may add punctuation, timing cues, emphasis, emotion, and physical reactions, but it should not quietly rewrite the claims or change the spoken words.
Input: I stopped buying energy drinks that made me crash by lunch. Output: [confessional, slightly breathless] I stopped buying energy drinks—[brief pause] that made me crash by lunch.
This makes the supervision legible. A reviewer can compare input and output, verify wording preservation, and decide whether the direction is supported by the audio rather than invented from the sentence alone.
02 / Inkling
Why an audio-language model belongs here
A text-only model can guess how a sentence might be performed. Inkling can accept audio as evidence. That matters for features such as a smile in the voice, a swallowed laugh, a sharp inhalation, a deliberately awkward pause, or the change from conversational proof to an urgent close.
Prosody
Pitch movement, intensity, stress, and pace can be described rather than guessed.
Reaction
Laughter, breath, strain, hesitation, and smiling delivery can carry persuasive meaning.
Not an audio generator
Inkling interprets audio and generates text. The speech renderer is a separate model with a separate job.
03 / LoRA
A small trainable layer over a fixed base
LoRA fine-tuning does not rewrite every parameter in Inkling. The base model remains fixed while a comparatively small adapter learns task-specific updates. That adapter can be loaded for inference or continued from a training-state checkpoint.
FixedBase Inkling
General audio-and-language capability remains the foundation.
UpdatedRank-32 adapter
Supervised loss adjusts the narrow expressive-direction layer.
ServeSampler checkpoint
Loads base plus adapter to generate responses.
ResumeTraining state
Preserves optimizer state when additional training is intentional.
The adapter is remote. A local dashboard can call it, but the model weights were not downloaded into this publication and are not exposed by the public site.
04 / Two objectives
Ad writing and performance direction are different skills
The project initially blurred two outcomes: writing a strong new ad from a product brief, and adding expressive direction to an already approved script. The training pairs mostly support the second.
| Objective | Input | Output | Best evaluation |
|---|---|---|---|
| Ad writing | Product, audience, facts, offer | New UGC script | Creative quality and claim accuracy |
| Expressive direction | Locked script, optional audio/context | Same words plus performance cues | Wording preservation and blind listening |
A future system may chain both tasks. Training and evaluating them separately first makes failures understandable.