---
name: Audio Processor - Segment-First Architecture
description: Revised pipeline architecture - detect breaks and split into segments BEFORE transcription for complete content capture
type: project
---

## Revised Pipeline Architecture (decided 2026-03-22)

Shows are almost always 4 segments per hour (8 total for a 2-hour show). Extra breaks are rare.

**Old approach:** Transcribe full episode -> truncate to fit LLM context -> analyze (loses content)

**New approach:** Detect breaks first (audio-only) -> split into ~8 segments -> transcribe each -> analyze each with full context -> cross-segment synthesis

### Pipeline Order

1. **Audio-level break detection** (no transcript needed) — loudness/compression jumps, silence gaps, known bumper fingerprints, HR1/HR2 boundary
2. **Split into segments** — ~7-15 min each, complete audio chunks
3. **Transcribe each segment** — smaller files, complete content, no truncation
4. **Analyze each segment** — full transcript fits in LLM context window easily
5. **Cross-segment synthesis** — detect topics spanning segments, callbacks ("going back to what we said before the break"), narrative arc
6. **Generate content** — blog posts, forum posts, episode summary from complete analysis

### Key Insights

- 4 segments/hour is a strong structural prior for break detection — if 12-18 min into a segment and audio signatures appear, almost certainly a break. At 5 min, probably not.
- Each segment transcript is ~5-10K chars — fits in any LLM context with room for detailed prompts
- Cross-segment synthesis pass is new and essential for catching callbacks and recurring topics

**Why:** Solves the context window truncation problem that loses show content. Each segment gets complete analysis.

**How to apply:** This is the architecture direction for all future audio processor work. The existing Stage 3 segment detector needs to work without transcript input (audio-only signals). Stage 6 analyzer needs per-segment + synthesis passes.