Voice Transcription Application

A real-time voice transcription application with spectrogram visualization, built using Python and OpenAI's Whisper model. The application provides real-time transcription of speech with adjustable parameters and supports multiple languages.

Features

• Real-time voice transcription using OpenAI's Whisper model
• Live spectrogram visualization
• Support for multiple languages
• Adjustable audio chunk sizes
• Silence threshold control
• GPU acceleration support
• Multiple Whisper model options (tiny, base, small, medium, large)
• Configurable minimum and maximum speech duration
• Input device selection
• Real-time transcription display with ordered results

Requirements

• Python 3.11 or higher
• CUDA-capable GPU (optional, for better performance)
• Microphone input device

Installation

1. Clone the repository:

git clone <repository-url>
cd VoiceChat

2. Create and activate a virtual environment:

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

3. Install dependencies:

pip install -r requirements.txt

Usage

1. Run the application:

python voice_recorder.py

2. Configure the application:

   • Select your input device from the dropdown
   • Choose your preferred language
   • Select the Whisper model size (tiny, base, small, medium, large)
   • Adjust minimum and maximum chunk lengths using the sliders
   • Set the silence threshold to control speech detection sensitivity

3. Start recording:

   • Click the "Start" button or press the space bar
   • Speak into your microphone
   • View real-time transcription in the text box
   • Click "Stop" or press space bar again to stop recording

Controls

• Input Device: Select your microphone
• Language: Choose the language for transcription
• Whisper Model: Select the model size (affects accuracy and performance)
• Min Chunk Length: Minimum duration of speech segments (0.1-1.0 seconds)
• Max Chunk Length: Maximum duration of speech segments (1.0-10.0 seconds)
• Silence Threshold: Control speech detection sensitivity (0.001-0.1)
• Start/Stop: Toggle recording

Performance Tuning

1. Whisper Model Selection:

   • Tiny: Fastest, lowest accuracy
   • Base: Good balance of speed and accuracy
   • Small: Better accuracy, moderate speed
   • Medium: High accuracy, slower
   • Large: Best accuracy, slowest

2. Chunk Length Settings:

   • Smaller chunks: More responsive but may miss context
   • Larger chunks: Better context but higher latency
   • Adjust based on your speaking style and system performance

3. Silence Threshold:

   • Lower values (0.001-0.01): More sensitive to quiet sounds
   • Higher values (0.01-0.1): Require louder speech
   • Adjust based on your microphone and environment noise

4. GPU Usage:

   • Enable CUDA for better performance
   • Monitor GPU memory usage with larger models
   • Adjust chunk sizes if experiencing memory issues

Supported Languages

The application supports transcription in multiple languages including:

• English
• Spanish
• French
• German
• Italian
• Portuguese
• Dutch
• Russian
• Japanese
• Korean
• Chinese
• Arabic
• And many more...

Troubleshooting

1. No Audio Input:

   • Check if your microphone is properly connected
   • Verify the selected input device in the dropdown
   • Adjust the silence threshold if needed

2. Poor Transcription Quality:

   • Try using a larger Whisper model
   • Adjust the chunk length settings
   • Ensure clear audio input
   • Fine-tune the silence threshold

3. Performance Issues:

   • Switch to a smaller Whisper model
   • Increase chunk lengths
   • Check GPU memory usage
   • Adjust silence threshold to reduce processing

4. Memory Issues:

   • Use a smaller Whisper model
   • Increase chunk sizes
   • Enable CUDA memory optimization
   • Clear GPU cache regularly

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• OpenAI for the Whisper model
• SoundDevice for audio input handling
• VisPy for spectrogram visualization
• PyTorch for deep learning support