Advanced Options
Models
All MLX-compatible Whisper models from Hugging Face are supported.
Available Models
| Model | Size | Accuracy | Speed | RAM |
|---|---|---|---|---|
whisper-large-v3-mlx (default) |
~3GB | Excellent | Slow | ~8GB |
whisper-large-v2-mlx |
~3GB | Excellent | Slow | ~8GB |
whisper-medium-mlx |
~1.5GB | Very Good | Medium | ~4GB |
whisper-small-mlx |
~500MB | Good | Fast | ~2GB |
whisper-base-mlx |
~150MB | Fair | Very Fast | ~1GB |
whisper-tiny-mlx |
~75MB | Basic | Fastest | ~512MB |
All models prefixed with mlx-community/
When to Use Each Model
- Large: Accented/technical content, maximum accuracy
- Medium: General purpose, good balance
- Small: Clear audio, need speed
- Base/Tiny: Quick drafts, limited resources
Batch Processing
Process multiple files:
# All files in directory
for file in audio/*.mp3; do
macscribe "$file" --output "transcripts/$(basename "$file" .mp3).txt"
done
# Save to directory with auto-generated names
for file in *.mp3; do
macscribe "$file" --output transcripts/
done
Long Content
For videos >2 hours, use a faster model:
Workflows
Transcription Pipeline
#!/bin/bash
INPUT="$1"
OUTPUT="$2"
MODEL="${3:-mlx-community/whisper-large-v3-mlx}"
macscribe "$INPUT" --model "$MODEL" --output "$OUTPUT"
Text Processing
# Word count
macscribe audio.mp3 --output transcript.txt
wc -w transcript.txt
# Search keywords
macscribe video.mp4 --output transcript.txt
grep -i "keyword" transcript.txt
Version Control
macscribe meeting.mp3 --output transcripts/2024-01-01.txt
git add transcripts/2024-01-01.txt
git commit -m "Add meeting transcript"
Optimization
Speed: Use smaller models
Accuracy: Use default (large) model, ensure high-quality audio
Shell Aliases
Add to .bashrc or .zshrc:
# Fast transcription
alias qscribe='macscribe --model mlx-community/whisper-medium-mlx'
# Auto-save
tsave() {
macscribe "$1" --output "${1%.*}.txt"
}
Troubleshooting
URLs with special characters: Use quotes
Corrupted files: Convert first