CORIS.THIRI.AI / EXPERIMENTS / WIDGET 01

Scribe Word Slicing vs Silence Detection Trap

Standard AI Failure Mode
Naive Silence Detector (-40 dBFS)

Cuts at inter-sentence pause (0.43s), mistaking breath for end-of-speech. Clips initial "D" on second sentence.

THIRI Grounded Architecture
Scribe Word Boundary (-0.20s Grid Snap)

Extracts Scribe word timestamp ($t = 1.85\text{s}$), sets cut at $t = 1.65\text{s}$, and snaps boundary to 24fps grid.

AUDIO TIMELINE: "In a sentimental mood... Duke Ellington was right." CUT POINT: 1.82s (FALSE CUT)
Pre-Roll Offset: -0.20s Lead time before first word consonant transient
Silence Threshold: -40 dBFS Energy cutoff triggering false break detector
Video Frame Grid: Snaps audio cuts to exact video frame boundaries
Dropped Phonemes 1 CLIPPED
Frame Jitter 38.4 ms
Breath Headroom 0.00 ms
Cut Status UNALIGNED
👤 Non-Tech / Creator Experience Click Any Word to Audition
Interactive Karaoke Word Rescue

Everyday Analogy: Standard silence cutters act like an impatient automatic door that slams shut the moment you pause to take a breath, chopping off the first letter of your next sentence ("...uke Ellington" instead of "Duke Ellington"). Click the words below to hear and see the difference.

⚠️ GUILLOTINE ACTIVE: The "D" in "Duke" is chopped off by the silence cutter!
💻 Developer / Live SDK Inspector
Scribe Timestamp + Frame-Grid Snapper

Live-computed FFmpeg trim filter and Scribe word-boundary math updating dynamically from your sliders above: