Speech-to-Text Productivity Tool for Windows

The bandwidth in speech is much higher than typing, this is a Windows desktop application that enhances productivity by converting spoken words into text. The application leverages the Faster Whisper implementation of the Whisper model for high-performance speech recognition, with all processing performed locally on your device to ensure data privacy.

Features

• Real-time speech-to-text transcription with up to 4x faster processing than normal whisper
• Two input modes:
  • Continuous Mode: Automatically detects speech using VAD
• Local processing with no data leaving the device
• Inserts transcribed text into the focused text box
• Transcription history with export capability
• Optimized GPU inference with 8-bit quantization support

System Requirements

• Operating System: Windows 10/11
• Hardware:
  • CPU: Modern multi-core processor (Intel Core i5/i7 or AMD Ryzen 5/7 or equivalent)
  • RAM: Minimum 8GB
  • GPU: (Optional) NVIDIA GPU with CUDA support (minimum 4GB VRAM, 8GB recommended)
  • Microphone: Any working microphone (built-in or external)
• Software:
  • Python 3.12+ with pip
  • CUDA 12.1

Installation

1. Clone the repository:

   git clone https://github.com/yourusername/speech-to-text-tool.git
   cd speech-to-text-tool
   

2. Install the required dependencies:

   pip install -r requirements.txt
   

3. Download the Faster Whisper model:

   # Create models directory
   mkdir -p ~/whisper-models
   
   # Authenticate to HuggingFace using the CLI login tool
   Run "huggingface-cli login" to authenticate 
   
   
   # Download the model (choose one):
   # For best accuracy (larger model, but slower on CPU):
   python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='distil-whisper/distil-large-v3', local_dir='~/whisper-models/distil-large-v3', local_dir_use_symlinks=False)"
   
   

4. Run the application:

   # CPU mode (default)
   python run.py
   
   # Or with GPU acceleration if you have CUDA 12.x installed (highly recommended):
   python run.py  --use-cuda
   

GPU Acceleration Setup

Note: The application runs in CPU mode by default for maximum compatibility. To use GPU acceleration for faster transcription:

1. Install NVIDIA CUDA Dependencies:

   • Download and install NVIDIA CUDA Toolkit 12.1 (required for latest faster-whisper)
   • Download and install NVIDIA cuDNN compatible with your CUDA version
   • Add CUDA bin directory to your system PATH environment variable

2. Run with the GPU flag:

   python run.py --use-cuda
   

3. Verify GPU usage:

   • The application will display a message confirming CUDA acceleration is enabled
   • You should see GPU memory usage in Task Manager or nvidia-smi
   • Transcription should be significantly faster than CPU mode

Usage

1. Start the application:

   python run.py --use-cuda
   

2. Configure the model:

   • Open Settings
   • Set the model directory to your downloaded model path

3. Select your microphone from the dropdown menu.

4. Click "Start Listening" to begin.

5. Focus on the text field where you want the transcribed text to appear.

6. Using Continuous Mode:

   • Simply speak naturally
   • Pause briefly between sentences
   • The application will automatically detect speech, transcribe it, and insert text

7. Click "Stop" when you're finished with all transcription.

Command-line Arguments

The run.py script supports several command-line arguments:

• --model-path PATH: Specify the path to your Whisper model directory
• --use-cuda: Enable CUDA GPU acceleration (requires CUDA 12.x)
• --debug: Enable debug logging
• --test: Run system tests
• --performance: Run performance tests

Building into a Standalone Windows Executable

Work in progress, will probably be in the next commit

To Do :

• Fix Build process and run from standalone executable
• Use newer Whisper-V3 Turbo model
• Increase accessibility for smaller GPU using smaller models
• Experiment with LLM-aided correction for inaccurate transcriptions
• Add the ability to dictate commands, 'close this windows' , 'press the maximize button' , etc....

Testing

Run the comprehensive test suite:

# Run system tests
python run.py --test --model-path ~/whisper-models/distil-large-v3

# Run performance tests
python run.py --performance --model-path ~/whisper-models/distil-large-v3

Troubleshooting

Performance Optimization

• Reduce Latency:

  • Run in debug mode (--debug) to see detailed performance traces
  • Place your cursor in the target field before speaking
  • Speak clearly with natural pauses between sentences
  • For best experience, use a high-quality microphone in a quiet environment

• CPU Mode (Default):

  • Use smaller model sizes for faster processing (small or medium)
  • Increase the CPU threads in settings if you have a multi-core processor
  • Enable VAD filter to reduce processing of non-speech audio
  • Close other CPU-intensive applications while using the app

• GPU Mode (If enabled):

  • Try using int8_float16 compute type if you have limited VRAM
  • Use a smaller model if you're experiencing out-of-memory errors
  • For older GPUs, install compatible ctranslate2:

    pip install --force-reinstall ctranslate2==3.24.0
    

CUDA Issues and Fix History

The application currently defaults to CPU mode due to incompatibilities between faster-whisper library dependencies and various CUDA versions. Specifically, newer versions of faster-whisper require:

• CUDA 12.1 libraries (specifically cublas64_12.dll)
• CTranslate2 compiled against CUDA 12.1
• Matching cuDNN libraries

If you're experiencing the error Library cublas64_12.dll is not found or cannot be loaded when trying to use GPU mode, you have the following options:

1. Update to CUDA 12.1: Install the latest CUDA 12.x toolkit and compatible cuDNN
2. Use CPU mode: Our app now defaults to CPU mode for reliability

Model Loading Issues

• Check model path in settings is correct
• Verify model files are properly downloaded
• Try downloading a different model size
• Check disk space is sufficient

Voice Activity Detection Issues

• Adjust VAD sensitivity in settings
• Move closer to the microphone
• Ensure microphone is properly selected and working
• Test in a quiet environment first

Privacy and Security

This application processes all audio data locally on your device using Faster Whisper. No data is sent to external servers, ensuring your privacy and security.

License

To do whatever you want with it license

Acknowledgements

• Faster Whisper for the optimized Whisper implementation
• CTranslate2 for the fast inference engine
• Silero VAD for voice activity detection
• PyQt for the graphical user interface
• PyWinAuto for Windows automation