whisper-batch

Batch transcribe all audio files in a directory using faster-whisper on GPU.

Setup

pip install faster-whisper
pip install python-docx   # optional, for .docx output
pip install nvidia-ml-py  # optional, for VRAM-based auto model selection

Requires CUDA-capable GPU(s). Falls back to CPU automatically if none detected.

Usage

# Transcribe all audio files in a directory (auto-detects GPU + best model)
python transcribe.py /path/to/audio/files

# Use a specific model
python transcribe.py /path/to/audio/files --model large-v3

# Output to a different directory
python transcribe.py /path/to/audio/files --output /path/to/transcripts

# Include timestamps
python transcribe.py /path/to/audio/files --timestamps

# Word document output (also generates .txt companion)
python transcribe.py /path/to/audio/files --format docx --timestamps

# SRT subtitle format
python transcribe.py /path/to/audio/files --format srt

# Search subdirectories recursively
python transcribe.py /path/to/audio/files --recursive

# Force CPU
python transcribe.py /path/to/audio/files --device cpu

Auto-detection

When --model auto (the default), the script detects free VRAM and picks the best model:

Free VRAM	Model	Compute
>= 6 GB	large-v3	float16
>= 3.5 GB	large-v3	int8
>= 2 GB	medium	int8
>= 1 GB	small	int8
CPU	small	int8

Supported Audio Formats

mp3, wav, m4a, flac, ogg, wma, aac, opus, webm, mp4, mkv, avi, mov