Timecoder

YouTube Transcript Analyzer with AI-Powered Semantic Segmentation

Timecoder is a Python application that analyzes YouTube video transcripts and organizes them into semantically related segments with timestamps. It's a demonstration project showcasing integration of modern NLP libraries for practical text analysis tasks.

What makes it interesting?

• Semantic Analysis: Uses SentenceTransformer embeddings to group transcript segments by meaning rather than arbitrary time intervals
• Multiple AI Models: Combines different AI approaches - KeyBERT for keywords, DistilBART for summaries, and local Gemma3:12b for text improvement
• Local Processing: Runs entirely on your machine without sending data to external services
• Real-world Application: Solves the actual problem of navigating long YouTube videos

Who is it for?

• Students learning NLP: Good example of combining multiple ML libraries
• Developers exploring AI integration in desktop applications
• Content creators who want to experiment with transcript analysis
• Anyone curious about semantic text processing

Why does it exist?

This started as an experiment to see if AI could automatically create better structure for YouTube's often messy auto-generated transcripts. It's primarily a learning project that demonstrates practical applications of transformer models and semantic similarity.

Tech Stack

Core Technologies:

• Python 3.8+ - Main language
• SentenceTransformer - Text embeddings for semantic similarity
• KeyBERT - Keyword extraction
• Transformers/DistilBART - Text summarization
• Ollama + Gemma3:12b - Local LLM for text post-processing

Supporting Libraries:

• NLTK - Text preprocessing
• YouTube Transcript API - Getting video transcripts
• Tkinter + ttkbootstrap - Desktop GUI
• tkinterweb - HTML rendering in GUI
• Requests - HTTP communication

Demo

Installation & Setup

Prerequisites

• Python 3.8+
• ~11GB disk space for the Gemma3:12b model
• Internet connection for initial setup

Installation Steps

1. Clone and setup

   git clone https://github.com/KazKozDev/timecoder.git
   cd timecoder
   python -m venv venv
   source venv/bin/activate  # or venv\Scripts\activate on Windows
   pip install -r requirements.txt

2. Install Ollama and model

   # Install Ollama (see https://ollama.com for platform-specific instructions)
   ollama serve
   ollama pull gemma3:12b

3. Setup NLTK data

   python -c "import nltk; nltk.download('punkt'); nltk.download('stopwords')"

4. Update Hugging Face token Edit line 27 in timecoder.py and replace with your token (or remove the login line)

Quick Start

python timecoder.py

Usage

1. Launch the application
2. Paste a YouTube URL (must have available transcripts)
3. Click "Analyze Transcript" and wait
4. View the timestamped segments in the output
5. Use "Copy All" to get the formatted text

Note: Processing can take several minutes depending on video length and your hardware.

How It Works

Processing Pipeline

1. Extract transcript from YouTube using their API
2. Clean and preprocess text (remove filler words, normalize)
3. Create chunks of transcript segments for analysis
4. Generate embeddings using SentenceTransformer
5. Group by similarity using cosine similarity thresholds
6. Extract topics with KeyBERT or DistilBART
7. Post-process with Gemma3:12b for better formatting
8. Display results in GUI with timestamps

Key Implementation Details

• Uses dynamic similarity thresholds to balance segment granularity
• Falls back to forced segmentation if semantic grouping produces too few segments
• Includes comprehensive error handling for flaky AI model responses
• Processes everything locally for privacy

Limitations & Known Issues

• Processing Time: Can be slow on older hardware (several minutes per hour of video)
• Memory Usage: Requires significant RAM during processing (~2-4GB)
• Model Dependencies: Relies on external models that may not always be available
• No Persistence: Results aren't saved automatically
• Limited Error Recovery: May fail on videos with unusual transcript formats
• No Testing: This is a demo project without formal test coverage

Project Structure

timecoder/
├── timecoder.py          # Main application
├── requirements.txt      # Python dependencies
├── README.md            # This file
├── LICENSE              # MIT license
└── .gitignore          # Git ignore rules

Contributing

Feel free to fork and experiment! This is a learning project, so I welcome:

• Bug fixes
• Performance improvements
• UI enhancements
• Better error handling

No formal contribution guidelines - just open an issue or PR.

---

If you like this project, please give it a star ⭐

For questions, feedback, or support, reach out to:

Artem KK | MIT LICENSE

Disclaimer: This is a learning/demo project, not production software. Use at your own risk.

A practical exploration of NLP and semantic analysis