aninews ep 18 — switched to edge tts, whisper-aligned captions, and a people-story rule

spacex's record IPO shipped — biggest IPO in history, and one share now buys a rocket, starlink, and a chatbot that loses six billion a year. but the real work this episode was the pipeline, not the script.

switched narration off google's chirp3-hd to edge TTS (andrew). free, no key, and it just sounds better. the catch: edge has no SSML, so the punchline break-pauses are gone. pacing comes from punctuation now.

the bigger fix was captions. google's voice silently strips the word-timing marks, so captions were estimated and drifted off the audio. now the stitch does forced alignment — whisper transcribes the rendered narration back into timed words (base.en, in a separate py3.12 venv since torch has no py3.14 wheels yet), aligns them to the script, and the captions land on the word. provider-agnostic, so it works no matter which voice made the audio.

the editorial lesson cost the most. first cut of beats 3-4-5 was a tour of an engine, a pump, a furnace. no people in frame. boring. the rule that came out of it: a story about things has to be told as a story about people — cast the real figures the news is actually about, not faceless extras or some invented mascot to pad the count. re-cut those beats around musk physically working the machine. better.

~$16 in generation, one daily-limit wall mid-render, 72 seconds out.