YouTube Transcript Analysis

A Python tool that improves YouTube's auto-generated transcripts using DeepSeek LLM and compares these with high-quality Whisper transcriptions using semantic similarity with sentence-BERT model. The project's goal is to create high-quality transcripts that aren't as computationally expensive as Whisper.

Results

Aggregate Similarity Distribution

This plot shows the distribution of sentence-level similarities between Whisper and YouTube transcripts. The blue line represents the original YouTube transcripts (mean: 0.241), while the orange line represents the cleaned transcripts (mean: 0.761). Higher similarity scores indicate better alignment with the Whisper transcription. The average improvement is 39.99% for n=1811 sentences and 1,659 tokens per minute of video on average.

Individual Analysis

Analysis for video sGUjmyfof4Q:

• Original similarity: 17.64%
• Cleaned similarity: 65.26%
• Improvement: 47.61%
• Total tokens used: 27,018

Analysis for video sNa_uiqSlJo:

• Original similarity: 25.42%
• Cleaned similarity: 81.78%
• Improvement: 56.35%
• Total tokens used: 38,903

Analysis for video x9Ekl9Izd38*:

• Original similarity: 86.81%
• Cleaned similarity: 71.85%
• Improvement: -14.96%
• Total tokens used: 14,912

Analysis for video w9WE1aOPjHc:

• Original similarity: 22.47%
• Cleaned similarity: 81.25%
• Improvement: 58.78%
• Total tokens used: 136,026

Analysis for video CSE77wAdDLg:

• Original similarity: 19.68%
• Cleaned similarity: 71.83%
• Improvement: 52.15%
• Total tokens used: 41,579

Notes:

• *It seems some youtube transcripts are significantly better than others.
• The sample size is probably too small.
• Similarity distributions for individual videos are in the plots/ folder.
• I need to see how big I can make the chunks before quality drops off to reduce API calls.

Features

• Downloads audio from YouTube videos
• Generates high-quality transcriptions using OpenAI's Whisper model
• Fetches YouTube's auto-generated transcripts
• Chunks and iteratively improves transcript quality using DeepSeek LLM
• Compares transcripts using semantic similarity analysis with SBERT
• Generates individual similarity distribution plots for each video
• Creates aggregate similarity distribution plots across all processed videos
• Supports batch processing of multiple videos
• Caches results to avoid reprocessing

Requirements

• Python 3.8+
• FFmpeg (for audio processing)
• CUDA-compatible GPU (optional, for faster transcription)

Installation

1. Clone the repository:

git clone https://github.com/brickbrycebrick/transcript_cleaner.git
cd youtube_summary

2. Create and activate a virtual environment:

python -m venv env
source env/bin/activate  # On Windows: env\Scripts\activate

3. Install dependencies:

pip install -r requirements.txt

4. Create a .env file with your API keys:

DEEPSEEK_API_KEY=your_api_key_here

Project Structure

youtube_summary/
├── data/
│   ├── audios/                # Downloaded audio files
│   ├── transcriptions/        # Whisper transcriptions
│   └── summary_transcripts/   # Processed YouTube transcripts
├── plots/                     # Similarity distribution plots
├── ytsummary/
│   ├── analysis.py           # Transcript comparison and visualization
│   ├── cleaner.py           # Transcript cleaning using AI
│   ├── yt_summary.py        # YouTube transcript processing
│   └── yt_transcriber.py    # Whisper transcription logic
├── run.py                    # Main pipeline orchestration
├── .env                      # API keys and configuration
├── .gitignore               # Git ignore rules
└── requirements.txt         # Python dependencies

Usage

1. Process videos through the entire pipeline:

from run import Pipeline
import asyncio

async def main():
    video_urls = [
        "https://www.youtube.com/watch?v=example1",
        "https://www.youtube.com/watch?v=example2"
    ]
    
    pipeline = Pipeline()
    results = await pipeline.process_videos(video_urls)

if __name__ == "__main__":
    asyncio.run(main())

This will:

• Check for existing transcripts and process missing ones
• Generate similarity analysis for each video
• Create individual similarity distribution plots
• Generate an aggregate similarity plot
• Print summary statistics

2. Run analysis only:

from ytsummary.analysis import TranscriptAnalyzer

analyzer = TranscriptAnalyzer()
results = analyzer.analyze_transcripts()

3. Just transcribe videos:

from ytsummary.yt_transcriber import YouTubeTranscriber
import asyncio

async def main():
    transcriber = YouTubeTranscriber()
    results = await transcriber.transcribe_videos([
        "https://www.youtube.com/watch?v=example"
    ])

if __name__ == "__main__":
    asyncio.run(main())

Output Files

The tool generates several types of output files:

• data/summary_transcripts/{video_id}.json: Contains original and cleaned YouTube transcripts
• data/transcriptions/{video_id}.json: Contains Whisper transcriptions
• plots/similarity_distribution_{video_id}.png: Individual video similarity plots
• plots/aggregate_similarity_distribution.png: Combined similarity analysis across all videos

Notes

• Transcriptions are cached in their respective directories
• Audio files are saved in the data/audios directory
• The tool skips processing if a transcription already exists
• GPU acceleration is used automatically if available for Whisper transcription
• Similarity plots show distribution of sentence-level similarities
• The aggregate plot combines data from all processed videos