Audience
Speech teams, multimodal researchers, model evaluators
Chris Royse field notes
Full ICL Carries The How, Not Just The Who
Full in-context learning gives the voice model waveform, transcript, cadence, mic character, and delivery style, not only an identity vector.
Media / alldata.md / 54:47
Core idea
Speaker embeddings can capture identity while still losing accent, cadence, room tone, and style. Full ICL preserves more of the signal.
Voice Cloning
Watch on YouTube· 54:47
Full ICL Carries The How, Not Just The Who
If a lab cares about controllable generation, the conditioning path matters as much as the backbone.
Watch videoOpen the full video on YouTubeWhat to take from it
The videos are raw build context. These notes translate them into the shortest useful frame for creators, companies, and AI lab readers.
Identity is not the same as delivery.
Reference selection affects the final clone.
Candidate scoring should be logged as provenance.
Continue this thread.
Related notes stay inside the same problem area first, then move to the next useful context.
Watch + read / 14:09
ClipCannon Turns Video Into Inspectable Data
ClipCannon breaks video into transcripts, frames, scenes, emotion, speaker, prosody, highlights, storyboards, and provenance.
Watch + read / 13:09
An AI-Native Editor Starts With Analysis
The editor only works because the system already knows scenes, transcript timing, narrative flow, captions, crops, and render constraints.
Watch + read / 7:49
The Meeting Bot Is A Multimodal Alignment Problem
A real-time avatar has to preserve voice, face, expression, timing, conversation state, and meeting latency all at once.
Make it concrete.
Send the audience, data type, target task, proof bar, and sharing limits.