Watch an amateur AI-edited talking-head video with the sound up, and you can spot naive FFmpeg silence removal within four seconds.
Every sentence ends with the speaker’s mouth frozen half-open for 300 milliseconds, or worse, the hard consonant at the end of a price quote gets chopped clean off because an aggressive decibel gate slammed shut too early.
Sentence-level SRT files cannot rescue that cut either. Knowing a sentence started at 00:04.100 and ended at 00:09.800 tells an editor nothing about the 1.1-second breath pause right in the middle where the founder lost their train of thought.
The cleanest fix costs zero API dollars and runs right on your laptop before your clip ever leaves disk: local Whisper word-level timestamps.
On an Apple Silicon Mac using mlx-whisper or whisper.cpp, transcribing sixty seconds of spoken audio with word-accurate { word, start, end } timings finishes in about three seconds locally. You pay nothing to external transcription APIs, and you keep raw rehearsal takes off third-party training logs.
When you open a talking_head project via create_project on https://mixi.mixbash.com/api/agent/mcp and upload your mixbash-intake-v1 MP4 with prepare_upload and complete_upload, you pass that word-timestamp array straight into register_media.
Inside Mixbash’s server harness, those millisecond word boundaries drive three things simultaneously: trimming false starts across multiple takes, tightening natural micro-pauses so the cadence feels energetic rather than robotic, and locking kinetic typography to exact spoken emphasis during generate_preview.
Because your client LLM (whether DeepSeek V4.1 Flash at ~$0.15/$0.60 per million tokens or Claude Haiku 5.5 at $0.10/$0.50) simply hands off the timestamp JSON and reads back short stage statuses, you never burn $20-to-$40 frontier output tokens asking a chat model to do floating-point timestamp subtraction in Python.
One grounded caveat from real client shoots: Whisper will occasionally misspell niche skincare ingredients, suburb names, or brand acronyms if someone recorded next to an espresso machine. Always scan the transcript text during draft_script or revise_script and fix proper nouns before running generate_preview and approve_review, because correcting a word early takes five seconds instead of a re-render.