testimonial-clips

Parse the VTT transcript for client-spoken positive keywords, cut 40-90 second segments with ffmpeg, re-transcribe with mlx_whisper for word timestamps, then burn karaoke-style ASS subtitles into vertical 1080x1920 video.

Generate an ASS subtitle file with per-word color override lines and \pos() positioning, then burn it using the ass= filter. Avoid force_style in the subtitles filter because commas break filter parsing; embed styles directly in the ASS file.

Zoom's built-in captions frequently mangle proper nouns and figures, so the workflow re-transcribes audio locally with mlx_whisper using the whisper-large-v3 model with word timestamps enabled for accurate subtitles.

Discard the video track and build a branded gradient quote card with a showwaves audio waveform overlaid via colorkey, keeping hook text, waveform, and subtitles in separate vertical layers.

Minimal ffmpeg builds lack libass, which is required for subtitle rendering. Split the pipeline by running transcription on one machine and rendering on a local machine with a full ffmpeg build.

Obtain per-clip client consent covering disclosure scope, real-name usage, and distribution channels, with revocation allowed anytime. Avoid embedding specific revenue figures in cards since transcripts may be inaccurate.