Real-Time YouTube Transcription with Whisper.cpp

This project enables real-time transcription of YouTube videos using your system's audio input. It captures audio through a loopback device and transcribes it.

Features

• Real-time audio capture from system audio output (loopback)
• Automatic transcription using Whisper.cpp models
• Support for multiple languages
• Timestamped transcription output
• Automatic saving of transcriptions to text files

Requirements

This project requires Python 3.12 and the following dependencies:

faster-whisper>=1.2.1
sounddevice>=0.5.3
numpy>=2.4.1

Install with:

pip install -r requirements.txt

Setup Instructions

1. Install Required Libraries

pip install faster-whisper sounddevice numpy

2. Download Whisper Model

The application requires a whisper.cpp model file. You can download one of the following models:

• Large model (recommended for accuracy):

  curl -L -o models/ggml-large-v3-turbo-q5_0.bin https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo-q5_0.bin

• Medium model (better balance of speed/accuracy):

  curl -L -o models/ggml-medium.en.bin https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.en.bin

• Base model (fastest, less accurate):

  curl -L -o models/ggml-base.en.bin https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin

3. Configure Audio Input

This application requires a loopback audio device to capture system audio. For Windows users, you'll need to:

1. Enable Stereo Mix (if available):

   • Open Windows Sound Settings
   • Go to "Recording" tab
   • Right-click on your default audio device and select "Properties"
   • In the "Listen" tab, check "Listen to this device"
   • Look for "Stereo Mix" in the list

2. Install Virtual Audio Cable (VB-Cable):

   • Download from vb-audio.com
   • Install and configure it as your loopback device
   • Set it as default recording device

Usage

Run the transcription script:

python transcribe.py

The program will:

1. List available audio devices
2. Prompt you to select a loopback device
3. Start capturing audio from the selected device
4. Transcribe audio in real-time with timestamps
5. Save output to a timestamped text file

Press Ctrl+C to stop transcription.

Output Format

Transcriptions are saved in text files with the following format:

[00:01:23] This is the transcribed text from the audio.
[00:01:45] Another sentence from the video.

Configuration Options

You can modify these settings at the top of transcribe.py:

• MODEL_PATH: Path to your whisper.cpp model file
• LANGUAGE: Language of the video (default: "en")
• SAMPLE_RATE: Audio sample rate (default: 16000)
• CHUNK_DURATION: Audio chunk duration in seconds before transcription (default: 5)

Troubleshooting

No Audio Devices Found

If no loopback devices are detected:

1. Ensure your audio drivers are up to date
2. Verify that Stereo Mix is enabled or VB-Cable is properly installed
3. Check that the selected device has input capabilities

Audio Quality Issues

• Make sure system volume is at a reasonable level
• Use high-quality audio sources for better transcription accuracy
• Ensure no other applications are using the audio device simultaneously

Notes

• This application works best with clear, well-recorded audio
• Transcription quality depends heavily on the audio source quality
• The application automatically resamples audio to 16000Hz for optimal performance with Whisper models

Dependencies

• Python 3.12+
• sounddevice - Audio input/output library
• numpy - Numerical computing library
• faster-whisper - Fast Whisper transcription library
• scipy - Scientific computing (included in requirements.txt)

Note: This application requires system audio capture capabilities. On Windows, this typically means having either Stereo Mix enabled or a virtual audio cable installed.