<aside> 🛠

Nudge Segment Cuts to Land on Word Boundaries — No More Mid-Word Slices

Video Analysis produces highlight segments with start/end times inferred by an LLM — but those timestamps often fall right in the middle of a spoken word. Snap Segments to Speech fixes that: it reads a word-timed transcript and nudges each segment's start_sec / end_sec to the nearest word edge (within a max distance you set). Feed the refined JSON into Video Trim and clips no longer start or end mid-syllable. A single Advanced control (max snap distance in seconds) keeps the whole thing under your control.

</aside>

What it does

Snap Segments to Speech is a JSON refinement tool that sits between Video Analysis and Video Trim in the highlight pipeline. It takes:

• Segments JSON from Video Analysis (with start_sec / end_sec per segment)
• Word-timed transcript from an upstream ASR / Speech-to-Text node

…and outputs a new segments JSON where each segment's boundaries have been snapped to the closest word edge in the transcript, up to a configurable maximum snap distance. The result: when the refined JSON drives Video Trim downstream, clips start and end on clean word boundaries — no mid-word audio, no half-cut consonants.

Problem it solves

• Prevent mid-word cuts – LLM-inferred segment boundaries are rarely word-aligned; without this step, Video Trim can produce clips that start halfway through a word
• Cleaner clip openings and endings – Trimmed clips sound intentional and complete when boundaries land on real speech events
• Better subtitle sync – Word-aligned boundaries pair cleanly with karaoke subtitles from Video Subtitle (Spoken) (no dangling partial words at clip edges)
• Preserve LLM intent – The tool only moves boundaries by a small tolerance (default 0.5 s); it doesn't rewrite the LLM's segmentation, just polishes the edges
• Deterministic post-processing – Same inputs + same tolerance = same output, so refinements are reproducible
• One-node insertion – Drop it between Video Analysis and Video Trim; no downstream changes needed
• Higher clip quality without more model calls – This is a data-refinement step, not another expensive LLM inference

Input/Output

<aside>

• Input: Segments JSON + word-timed transcript
  • Segments JSON (top input, orange port): Highlight-segment metadata from Video Analysis — one entry per segment with fields like start_sec, end_sec, title, hook, hook_sec, score
  • Transcript (bottom input, pink/magenta port): Word-timed transcript from an upstream ASR / Speech-to-Text node — each word carries a timestamp </aside>

<aside>

• Output: Refined segments JSON
  • Format: Same JSON shape as the input segments — one entry per segment, all original fields preserved
  • What changes: Only start_sec and end_sec are modified — each is nudged to the nearest word edge within the max-distance tolerance
  • What stays the same: All other fields (title, hook, hook_sec, score, etc.) pass through unchanged
  • Downstream: Feed directly into Video Trim — the physical cutting step now cuts on word boundaries </aside>

Configuration Options

Advanced Options