AI Video Clip Assistant

Problem

Editing long videos, live recaps, and talking-head clips often stalls at three points: inaccurate speech transcription, manual timeline decisions, and post-render review for black frames, clipping, long silences, and bad cuts. Reusing one source for social platforms also means repeated cropping, subtitle burn-in, and aspect-ratio adaptation, which can fragment into ad-hoc scripts.

How It Works

ai-video-clip turns editing into an auditable agent workflow: transcribe.py uses whisper, faster-whisper, or Chinese-optimized funclip to build a structured transcript; generate_edl.py turns that transcript plus a strategy into an EDL. Strategies include remove_fillers, extract_highlights, condense, topic_splits, social_clip, and custom. Each cut carries start/end/reason/confidence, so decisions can be reviewed, edited, and replayed. render_edl.py then executes the EDL with FFmpeg, applying audio crossfades, color correction, subtitle burn-in, and concat. self_eval.py checks duration, black frames, clipping, long silence, and cut boundaries, and can enter a bounded repair loop.

Fit and Limits

It fits local video files where you can describe the edit intent in natural language, especially Chinese filler removal, live highlight extraction, and batch social-format adaptation. It depends on FFmpeg and Python, and local ASR may use CPU/GPU resources; speaker-diarization options are optional. For strict brand review, complex color grading, or interactive NLE-style timeline work, treat the EDL as an intermediate artifact and route it into a more specialized editing pipeline.