Video Transcript Search Tool

This tool allows you to transcribe videos, create searchable embeddings, and perform semantic search through video transcripts. It uses OpenAI's Whisper for transcription and sentence transformers for semantic search capabilities.

Dependencies

Key libraries used:

• openai-whisper: For accurate video transcription
• sentence-transformers: For creating semantic embeddings
• faiss-cpu: For efficient similarity search
• torch: Required for ML models
• ffmpeg-python: For video processing
• Additional dependencies are listed in requirements.txt

Setup

1. Create and activate a virtual environment:

python3 -m venv venv
source venv/bin/activate  # On Windows use: venv\Scripts\activate

2. Install dependencies:

pip install -r requirements.txt

Project Structure

my_project/
├── videos/         # Place your .mp4 videos here
├── transcripts/    # Generated transcripts (JSON)
├── index/         # Generated FAISS index and metadata
├── src/           # Source code
└── run.py         # Main script

Usage

The tool provides three main commands:

1. Transcribe Videos

Place your .mp4 videos in the videos/ directory, then run:

python run.py transcribe

This will:

• Load the Whisper model
• Process all new videos in the videos/ directory
• Save transcripts as JSON files in transcripts/
• Skip already transcribed videos

2. Create Searchable Index

After transcription, create the search index:

python run.py embed

This will:

• Process all transcripts in transcripts/
• Create embeddings using sentence transformers
• Build a FAISS index for efficient similarity search
• Save the index and metadata in index/

3. Search Through Transcripts

To search through your video transcripts:

python run.py query

This will start an interactive query interface where you can:

• Type natural language queries to find relevant video segments
• See timestamps, video sources, and relevance scores
• Get quick access to specific moments in your videos

To exit the query interface:

• Type :q or :quit
• Or use Ctrl+C

Example query session:

> what did obama say about hope
[Results showing relevant transcript segments...]

> how many people were injured
[Results showing relevant transcript segments...]

Search Results Format

Each search result shows:

• 🎥 Source video file
• ⏱️ Timestamp (MM:SS format)
• 💬 Transcript text
• 📊 Relevance score

Performance Metrics

The tool provides timing breakdowns for searches:

• Embedding time: Converting query to vector
• FAISS search time: Finding similar segments
• Result formatting: Preparing output
• Total processing time

Notes

• Supports .mp4 video files
• Uses Whisper's base model (can be configured to small/medium)
• Provides semantic search (meaning-based, not just keyword matching)
• Maintains original video files untouched