Audio Transcription with Whisper

Transcribe audio files into text using OpenAI's Whisper speech recognition model.

Features

• Automatic model selection based on available GPU VRAM
• GPU acceleration (CUDA) with automatic CPU fallback
• Language auto-detection for multilingual audio
• Progress tracking with visual progress bar
• Memory efficient segment-by-segment processing
• Multiple audio formats supported (MP3, WAV, etc. via ffmpeg)

Requirements

Python Version

• Recommended: Python 3.9.9 (Whisper was trained and tested with this version)
• Supported: Python 3.8 - 3.11

Dependencies

pip install -r requirements.txt

Additional Requirements

• ffmpeg: Required for audio file loading
  • Download from ffmpeg.org or
  • Windows: choco install ffmpeg or scoop install ffmpeg
  • Mac: brew install ffmpeg
  • Linux: [you know better]

Recommended

• PyTorch: Version 1.10.1 (Whisper was tested with this version, but it also works with 2.8.0+cuda126)
• CUDA-Version: Install via PyTorch Get Started
• Audio file can be converted from mp4 file with (Windows):

  ffmpeg -i <video_filename>.mp4 -vn -acodec libmp3lame -q:a 0 <audoi_filename>.mp3

Usage

Basic Usage

python transcribe_audio.py --audio-file-path audio.mp3

Specify Model Size

python transcribe_audio.py --audio-file-path audio.mp3 --model-size medium

Custom Output Folder

python transcribe_audio.py --audio-file-path audio.mp3 --output-folder-path ./my_transcripts

Command Line Options

Option	Description	Default
--audio-file-path	Path to the audio file to transcribe	Required
--model-size	Whisper model (tiny, base, small, medium, large, auto)	auto
--output-folder-path	Folder where transcription will be saved	./transcripts

Model Sizes

Model	VRAM Required	Relative Speed	Accuracy
tiny	~1 GB	Fastest	Lowest
base	~1 GB	Very Fast	Low
small	~2 GB	Fast	Medium
medium	~5 GB	Moderate	High
large	~10 GB	Slow	Highest
auto	N/A	Auto-selected based on available VRAM	Varies

How It Works

1. Model Selection: Automatically selects the best Whisper model based on available GPU VRAM
2. Audio Loading: Loads and resamples audio to 16kHz (Whisper's expected format)
3. Language Detection: Detects the spoken language from the first 30 seconds
4. Segmented Processing: Processes audio in 30-second segments for memory efficiency
5. Transcription: Converts speech to text and saves to a .txt file

Output

• Transcription saved as: <audio_filename>.txt
• Located in the specified output folder (default: ./transcripts)
• Format: Plain text with spaces between segments

Troubleshooting

"CUDA not available"

• Script will automatically use CPU mode
• Install CUDA-compatible PyTorch for GPU support:

  pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126

"ffmpeg not found"

• Install ffmpeg and ensure it's in your system PATH
• Verify: ffmpeg -version

Incomplete transcription

• Check that your audio file is not corrupted
• Try a smaller model size if running out of memory
• Ensure audio file is in a supported format

License

This project uses OpenAI's Whisper model. Please refer to the Whisper repository for license information.