How This Works

This site is the output of an automated content pipeline purpose-built for Speak Your Truth — converting 8 years of YouTube livestream archives into a searchable, structured knowledge graph with per-segment metadata, clinical tagging, and speaker/theme backlinks.

In Plain Terms

Here’s what this system does, without the tech-speak:

Today, if SYT wants to find every clip where a speaker talks about anxiety, it takes hours of someone watching footage with a notepad. With this system, that’s a three-second filter.

Every time a video is posted, the system quietly:

1. Listens to the full video and transcribes every word — with precise timestamps, so any moment in any video can be jumped to directly.
2. Reads the transcript like an editor — identifying the distinct parts (a speaker’s story, a musical performance, a conversation) and tagging each one with themes (anxiety, grief, recovery, community healing…), people, events, and clinical categories from the CHIME-D framework that peer-support researchers use.
3. Publishes everything to this site — where every segment is its own page, cross-linked to every person, theme, and event it touches. The dots-and-lines map in the corner of each page is a visual index of those connections — click any node to jump to it, drag to rearrange, or hit the expand button for the full map.

The result is a searchable, structured archive of eight years of SYT’s work — one the content team can filter, reuse, and cite in minutes rather than hours, without needing technical skills. Every video’s audio is preserved offline on encrypted storage; every transcript is tracked in version control; every tag is traceable back to the exact moment in the video that justifies it. Nothing here is locked inside a single platform or vendor — if YouTube changes policy tomorrow, this archive keeps working.

The rest of this page is for the technically curious — a walkthrough of the stack, the stages, and the decisions behind them.

What’s Been Built

As of this build, the knowledge graph contains:

• 250 narrative segments — individually tagged, timestamped, and linked into the graph
• 113 people and organizations with dedicated hub pages
• 9 mental-health and community themes organizing the archive
• 8 SYT events spanning years of advocacy work
• CHIME-D clinical framework applied to every segment (Connectedness, Hope, Identity, Meaning, Empowerment, Difficulties) with supporting evidence quotes pulled directly from transcripts
• Four-layer taxonomy — format (the what), theme (the why), entity (the who), and event context (the where/when) — covering every segment
• Full git-tracked transcripts so the raw artifact of every video is preserved even if the LLM layer needs to be re-run

What This Unlocks for Speak Your Truth

The archive isn’t just a destination — it’s a substrate for everything else SYT does:

• Social media, on autopilot — Every segment carries its format, theme, and timestamps, so Instagram/TikTok managers can find clip-worthy moments by filter (e.g., “all spoken-word segments tagged theme:stigma_reduction under 90 seconds”) without re-watching hours of footage.
• Closed captioning for YouTube — The transcripts meet or beat YouTube’s auto-generated CC quality, especially for mental-health vocabulary, speaker names, and overlapping live-event audio. A small converter can emit .srt files ready for upload to every existing video — a significant accessibility win.
• Evidence-backed grant applications — The CHIME-D tagging means you can pull direct evidence quotes supporting any impact claim. No more anecdotal “we help people” — you can produce a funder-ready table: “N segments reference Connectedness, here’s the verbatim evidence.”
• Thematic content discovery — The graph view reveals unexpected connections between speakers, themes, and events. Which themes co-occur? Which speakers have touched on the most diverse topics? Surface these insights to guide programming.
• Research and academic partnerships — The structured, clinically-aligned metadata is exactly what peer-support researchers need. SYT’s archive becomes a citeable resource, not just a YouTube channel.
• Cross-platform repurposing — Newsletter blurbs, blog posts, podcast clip shows — anywhere SYT publishes, the archive can supply pre-tagged, searchable source material.
• Historical record — 8 years of peer advocacy, preserved in a format that outlives any single platform. If YouTube changes policy tomorrow, the transcripts and metadata are safe.

Built For, And With, Speak Your Truth

Everything on this site was produced by an automated pipeline running on commodity hardware — no cloud GPU costs, no per-minute transcription fees. The LLM calls are the only marginal cost, and the two-pass architecture keeps those bounded (segmentation uses a cheaper model; enrichment uses a stronger one for final quality).

The entire codebase is tracked in git, fully documented, and designed to be handed off — so SYT owns the substrate, not just the output.

The Pipeline

The system is designed around a “slow feed, fast chew” philosophy — downloads are rate-limited and patient (to avoid bot detection), but once audio is staged, transcription and metadata generation run in parallel across many workers.

Five stages, fully automated:

1. Download — yt-dlp pulls audio-only MP3s from the Speak Your Truth YouTube channel, with randomized sleep intervals and cookie support for age-restricted content.
2. Transcribe — Two parallel worker containers run OpenAI’s Whisper large-v3 model (via faster-whisper 1.1.0, quantized to int8 for speed) — producing word-level timestamps and segment text for every audio file.
3. Finalize — A two-pass LLM agent (Google gemini-2.5-flash-lite) reads the transcript, identifies narrative segments (Pass 1: segmentation) and enriches each with YAML metadata (Pass 2: themes, entities, CHIME-D clinical categorization, evidence quotes). The finalizer is resilient — retries failed API calls with exponential backoff and checkpoints progress so a crash doesn’t lose work.
4. Transform — A Python transformer explodes multi-segment outputs into individual pages, generates hub pages for every person/theme/event, and wires up the Obsidian-style wikilinks that drive the knowledge graph.
5. Publish — Quartz 4 renders the content into this site you’re reading now, with full-text search, backlinks, and the interactive graph view you see in the bottom-right corner.

Technology

Layer	Component	Version
Pipeline runtime	Python	3.11.15
Service orchestration	Docker Compose (5 services)	—
State & coordination	SQLite ledger	—
Audio extraction	yt-dlp	2026.03.17
Speech recognition	faster-whisper (large-v3, int8)	1.1.0
Whisper runtime	CTranslate2	4.4.0
Metadata LLM	Google gemini-2.5-flash-lite (via google-genai)	1.14.0
Static site	Quartz 4 (TypeScript/Obsidian-based)	—
Hosting (planned)	Cloudflare Pages + Cloudflare Access	—
Cold archive	LUKS-encrypted SSD + external ext4 drive	—

Every run writes a run-config.json capturing exact tool versions plus SHA-256 hashes of the taxonomy, entity registry, and finalizer prompt — so any output file can be traced back to the precise configuration that produced it. Reproducibility is a first-class concern.

---

Technology provided by Carbon Works LLC
Making Sense of Technology — straightforward technology guidance, community-rooted in Annapolis.