Skip to content
Mohsen Askar

Audio Transcription Module

Overview

The Audio Transcription module enables users to convert audio files to text using OpenAI's Whisper speech recognition technology. It forms the first step in the Audio2Topics workflow, providing accurate transcriptions of audio content for further text processing and topic modeling.

This module allows you to: - Load multiple audio files in various formats (MP3, WAV, M4A, FLAC) - Select from different Whisper model sizes based on your accuracy and speed requirements - Process files using either CPU or GPU acceleration - View and save transcriptions to text files

Main interface of the Transcriber Tab png png

Core Components

Transcriber Class

The Transcriber class serves as the main interface for audio transcription functionality. It manages the transcription process and provides methods for saving results and querying available resources.

Methods

Method Description Parameters Returns
transcribe_files() Starts transcribing audio files using a worker thread audio_files: Dict[str, bytes] - Dictionary mapping filenames to file content
model_name: str - Whisper model name to use		 Returns the worker thread that can be connected to signals
save_transcriptions() Saves transcriptions to text files transcriptions: Dict[str, str] - Dictionary with filenames as keys and transcriptions as values
output_dir: str - Directory to save files		 List[str] - List of paths to saved transcription files
get_available_devices() Gets list of available devices for transcription None List[str] - List of available devices (CPU and GPUs if available)
get_available_models() Gets list of available Whisper models None List[str] - List of available model names

TranscriberWorker Class

The TranscriberWorker class extends QThread to handle audio transcription in a background thread, keeping the UI responsive during processing.

Signals

Signal Description Parameters
progress_updated Emitted to update progress int: progress percentage, str: status message
transcription_completed Emitted when transcription is done dict: dictionary of transcriptions
error_occurred Emitted when an error occurs str: error message

User Interface

The TranscriberTab class provides the user interface for the audio transcription functionality.

UI Components

  • Audio Files Panel:
  • File list showing loaded audio files

  • Load and Clear buttons for file management

  • Transcription Controls:

  • Model selection dropdown (tiny, base, small, medium, large)
  • Device selection (CPU/GPU options)
  • Language auto-detection toggle
  • Transcribe button

  • Progress bar

  • Transcription View:

  • Text area showing the current transcription
  • Save button for exporting transcriptions

Usage Guide

Loading Audio Files

  1. Click the "Load Audio Files" button to open the file selection dialog
  2. Select one or more audio files (supported formats: MP3, WAV, M4A, FLAC)
  3. Selected files will appear in the file list on the left panel

Selecting Transcription Options

  1. Choose a Whisper model from the dropdown:
  2. tiny: Fastest, lowest accuracy (~75MB)
  3. base: Fast with improved accuracy (~142MB)
  4. small: Good balance of speed and accuracy (~466MB)
  5. medium: High quality for most use cases (~1.5GB)

  6. large: Maximum accuracy for challenging audio (~3GB)

  7. Select a device:

  8. CPU: Available on all systems

  9. CUDA: Available if you have an NVIDIA GPU with CUDA support

  10. Optionally toggle language auto-detection (enabled by default)

Starting Transcription

  1. Click the "Transcribe" button to begin processing
  2. The progress bar will display the current status
  3. Transcription speed depends on:
  4. Model size
  5. Audio file length
  6. Selected device (GPU is significantly faster)
  7. System specifications

Viewing and Saving Transcriptions

  1. After transcription completes, select a file from the list to view its transcription
  2. Click "Save Transcription" to export the current transcription to a text file
  3. Choose a location and filename in the save dialog
  4. The transcription will be saved as a plain text file

Whisper Models

Model Parameters Size Relative Speed Best For
tiny 39M ~75MB ~32x Quick transcriptions where accuracy is less critical
base 74M ~142MB ~16x Fast transcriptions with improved accuracy
small 244M ~466MB ~6x Good balance between speed and accuracy
medium 769M ~1.5GB ~2x High-quality transcriptions for most use cases
large 1550M ~3GB 1x Maximum accuracy for challenging audio

Language Support

Whisper supports transcription in numerous languages and can automatically detect the language being spoken. Best results are achieved with:

  • English
  • Norwegian

Tips for Better Results

  1. Audio Quality Matters:
  2. Use clear recordings with minimal background noise

  3. Higher quality audio files generally produce better transcriptions

  4. Model Selection:

  5. Start with the "medium" model for a good balance of accuracy and speed
  6. Use "turbo" for best balance between spped and accuracy
  7. Use "large" for difficult audio or when maximum accuracy is needed

  8. Use smaller models for faster processing or when resources are limited

  9. Processing Strategy:

  10. Split very long recordings into 10-30 minute segments for better results

  11. Test with a short sample to determine the best model for your specific audio

  12. Hardware Considerations:

  13. GPU acceleration dramatically improves processing speed
  14. Larger models require more memory (RAM)

Troubleshooting

Common Issues

Transcription fails to start - Check for active internet connection (required for initial model download) - Try a smaller model if experiencing memory issues - Ensure sufficient disk space for model download - Restart the application and try again

Poor transcription quality - Try a larger model (medium or large) - Check audio quality and consider pre-processing to reduce noise - For non-English content, verify language detection is working correctly - Split complex audio with multiple speakers into smaller segments

Slow transcription - Use a CPU with more cores or enable GPU acceleration if available - Try a smaller model (base or small) - Close other resource-intensive applications while transcribing - For long files, consider splitting them into smaller chunks

Out of memory error - Use a smaller model - Reduce batch size if transcribing multiple files - Close other applications to free up memory - Consider upgrading your system's RAM if this is a recurring issue