Resumex

OVERVIEW

Resumex takes a written story and returns a finished 1080×1920 MP4 with narration and synchronized captions. The whole pipeline runs on the machine: no API keys, no cloud services, no network calls in the default path.

THE PROBLEM

Producing short-form video from written material is mostly repetitive assembly — narrate, transcribe, time the captions, compose, encode — and the usual tooling pushes each of those steps through a different paid service.

APPROACH

Caption timing is read directly from the speech synthesis output instead of transcribing the audio back, which removes an entire error-prone stage. Rendering is a single deterministic FFmpeg pass, so the same input yields the same file. Completed work is tracked in SQLite, making batch runs resumable rather than repeated.

FEATURES

• Local text-to-speech with Kokoro, running on CPU (or a silent mode)
• Accepts .txt, .md, and .json stories; custom footage or generated gradient backgrounds
• Optional local Ollama models to draft titles or score a story before rendering
• Optional private upload to YouTube through browser-based OAuth

STACK

Python / FFmpeg / Kokoro / SQLite / Ollama