ToneLab
Our project generates four distinct, judge-ready captions for every fixed short video clip (30s–2min): formal, sarcastic, humorous-tech, and humorous-non-tech. Built on a Python + uv pipeline that calls Fireworks AI through its OpenAI-compatible API, it first extracts the audio and runs Whisper v3 Turbo for transcription (with automatic translation for non-English clips), then samples keyframes and describes them with a vision model. A captioner model then produces each style from a strictly grounded prompt that combines the transcript and the visual facts, keeping temperature low so captions never hallucinate events that aren't in the video maximizing the judge's accuracy score. Tone is enforced by four separate system prompts with few-shot examples, so each register stays unmistakable and free of style bleed maximizing the judge's tone score.
Category tags:
For Innovators & Creators