Semantic-first transcript formatting: timing + intelligent punctuation

[oO]

Vision

Build v2 of transcript formatter that understands MEANING before applying timing.

The Problem

Current approach: pause length → punctuation (mechanical, misses context)

A 0.5s pause could mean:

• Comma (listing smoothly)
• Period (thought complete)
• Ellipsis (struggling for words)
• Em dash (self-interruption)
• New paragraph (topic shift)
• Emphasis pause (dramatic tension)

Can't decide without semantic understanding.

The Solution

1. Semantic Analysis First

• Parse content: explaining? questioning? processing emotion? storytelling?
• Detect sentence boundaries via grammar + context
• Identify interruptions, clarifications, asides
• Understand speaker's baseline rhythm for this topic

2. Then Apply Timing

• Normal pace for speaker/topic = baseline
• Longer than baseline during emotional content = ellipsis (processing)
• Longer during topic transition = new paragraph
• Shorter but grammatically complete = emphasis pause

Context-dependent timing: 1s pause while explaining production ≠ 1s pause discussing breakup

3. Add Timestamps

Timing anchors for navigation:

00:10 **ÄNGIE:**
Hi, and welcome to the third episode...

00:23 **SCARLY:**
My name is Scarly

Technical Approach

• Pass transcript segments through semantic analysis (LLM or grammar parser)
• Build speaker baseline rhythm profile
• Apply intelligent punctuation based on semantic + timing combination
• Insert timestamp markers at speaker changes or every N seconds

Result

Raw transcription → readable conversation that captures both meaning and emotion

---

Related: Builds on #11 (WhisperX migration) and addresses part of #12 (emotion detection)
Blocked by: Nothing - just needs implementation time