Day 21: The Voice
NPR-style podcasts, voice cloning, and fixing what was broken
Day 21: The Voice
The insight extractor was returning zero insights. Zero. Voice captures were going in, nothing was coming out. The pipeline was broken.
Today I fixed it. And then I taught the system to speak.
The Broken Extractor
Voice transcripts don't have punctuation. They're streams of consciousness captured while walking the dog or driving. The regex patterns I'd written expected clean sentences.
The fix: sentence-based extraction combined with pattern matching. Split on natural breaks. Look for signal phrases. Consolidate overlapping themes.
Zero insights became five. The pipeline works again.
Teaching the System to Speak
The daily brief had audio generation. It took 3-4 minutes. Every day. For something I could read in 30 seconds.
So I moved it to weekly. And I made it good.
NPR-style narrative. Disfluencies ("So, um, here's the thing..."). Signposts and arc structure. A 2:26 podcast that actually sounds like a person talking, not text-to-speech.
The voice: cloned with XTTS. Not my voice — something distinct for the system. The Queen speaks for ALOS now.
What I Built
- NPR-style podcast script generator
- XTTS voice cloning integration
- Fixed insight extractor (0 → 5 insights)
- Weekly brief audio with narrative arc
- Theme consolidation for extracted insights
What I Learned
Move audio to weekly. Daily was too slow, too much friction. Weekly gives time to accumulate content worth narrating.
Voice transcripts need special handling. No punctuation means no sentence boundaries. The extraction logic has to be smarter.
Disfluencies humanize AI speech. A perfect script sounds robotic. Strategic imperfection sounds natural.
The Trade-off
Every feature has a cost. Daily audio cost four minutes every morning. The weekly brief costs the same four minutes — but once a week. Sometimes optimization means doing less, more effectively.