AI-Youtube-Shorts-Generator

Automatically downloads full-length YouTube videos using yt-dlp or similar library, storing them locally for subsequent processing. Handles authentication, format selection, and metadata extraction in a single operation, enabling offline processing without repeated network calls. The YoutubeDownloader component manages the download lifecycle and integrates with the transcription pipeline.

Converts video audio to text using OpenAI's Whisper model, generating word-level timestamps that map each transcribed segment back to specific video frames. The transcription output includes confidence scores and speaker diarization hints, enabling precise temporal mapping for highlight detection. Handles multiple audio formats and automatically extracts audio from video containers using FFmpeg.

Analyzes full video transcripts using GPT-4 to identify the most engaging, shareable segments based on content relevance, emotional impact, and audience appeal. The system sends the complete transcript to GPT-4 with a structured prompt requesting segment timestamps and engagement scores, then ranks results by predicted virality. This enables semantic understanding of content quality rather than simple keyword matching or silence detection.

Detects human faces in video frames using OpenCV with pre-trained Haar Cascade or DNN-based face detection models, then tracks face position and size across consecutive frames to maintain speaker focus during cropping. The system builds a spatial map of face locations throughout the video, enabling intelligent cropping that keeps speakers centered in the 9:16 vertical frame. Handles multiple faces and tracks the primary speaker based on face size and screen time.

Crops video segments from 16:9 (or other aspect ratios) to 9:16 vertical format while keeping detected speakers centered and in-frame. The system uses the face tracking data to calculate optimal crop windows that maximize speaker visibility while minimizing empty space. Applies smooth pan/zoom transitions between crop windows to avoid jarring frame shifts, and handles edge cases where speakers move outside the vertical frame boundary.

Combines multiple cropped video segments into a single output file, handling transitions, audio synchronization, and metadata preservation. The system uses FFmpeg's concat demuxer to join segments without re-encoding (when possible), applies fade transitions between clips, and ensures audio remains synchronized throughout. Supports adding intro/outro sequences, watermarks, and metadata tags for platform-specific optimization.

Coordinates the entire workflow from YouTube URL input to final vertical short output, managing state transitions between components, handling failures gracefully, and providing progress tracking. The main.py script implements a sequential pipeline that chains together download → transcription → highlight detection → face tracking → cropping → composition, with checkpointing to resume from failures. Includes logging, error recovery, and optional manual intervention points.

Exposes tunable parameters for each pipeline stage (highlight detection sensitivity, face detection confidence threshold, crop margin, transition duration, output resolution), enabling users to optimize for their specific content type and platform requirements. Configuration is managed through a JSON/YAML file or command-line arguments, with sensible defaults for common use cases (YouTube Shorts, TikTok, Instagram Reels). Supports platform-specific output presets that automatically adjust resolution, bitrate, and aspect ratio.

Enables processing multiple YouTube videos in sequence with a job queue, progress tracking, and optional parallelization. The system maintains a queue of URLs, processes them sequentially (or in parallel with worker threads), and provides real-time progress updates including estimated time remaining. Supports resuming interrupted batch jobs and generating summary reports of successes/failures.