For most Mac users, Whisper Base or Small provides the optimal balance of high speed and strong accuracy. For complex technical jargon or multilingual meetings, Whisper Large-v3 delivers the highest precision on Apple Silicon.
Balancing Speed, Precision, and System Resources
OpenAI released Whisper in several model sizes. Smaller models execute with blazing speed and tiny memory footprints, while larger models provide superior handling of heavy accents, obscure technical jargon, and noisy room environments.
Oyma lets you download and switch between models with a single click.
Model Tier Breakdown
| Model Tier | Parameters | Download Size | RAM Overhead | Relative Speed | Word Error Rate (WER) |
|---|---|---|---|---|---|
| Tiny | 39 M | 75 MB | ~250 MB | ~16x real-time | ~9.5% |
| Base | 74 M | 145 MB | ~400 MB | ~10x real-time | ~6.8% |
| Small | 244 M | 480 MB | ~900 MB | ~5x real-time | ~4.9% |
| Medium | 769 M | 1.5 GB | ~2.2 GB | ~2.5x real-time | ~3.8% |
| Large-v3 | 1,550 M | 3.1 GB | ~4.1 GB | ~1.5x real-time | ~2.9% |
Hardware Recommendations by Machine
- MacBook Air (M1/M2/M3 with 8 GB RAM):
- Recommended: Whisper Base or Small.
- Delivers rapid transcription while keeping memory free for other productivity apps.
- MacBook Pro (M1/M2/M3/M4 Pro or Max with 16 GB+ RAM):
- Recommended: Whisper Medium or Large-v3.
- Apple Silicon’s 16-core Neural Engine processes Large-v3 faster than real-time with virtually zero audible fan spin.
- Intel Macs:
- Recommended: Whisper Tiny or Base.
- Intel processors lack the Apple Neural Engine, so larger models run more slowly on x86 CPU cores.
Hardware-Specific Model Selection Matrix
To select the ideal Whisper model for your hardware:
- Whisper Base (145 MB): Best for general note-taking on MacBook Air and 8 GB RAM machines. 10x real-time speed.
- Whisper Small (480 MB): Excellent balance of high accuracy and low memory usage on 16 GB machines.
- Whisper Large-v3 (3.1 GB): Ultimate accuracy for heavy accents, noisy backgrounds, and multilingual audio on M-Pro/Max systems.
Effortless Model Switching Across Use Cases
You can download multiple Whisper model tiers and toggle between them depending on your current workflow: use Base for rapid daily journaling, and switch to Large-v3 for critical legal or academic interviews.
Model files are stored locally in your Application Support directory and can be deleted or re-downloaded at any time to manage disk storage according to your preferences.
Whisper Model Tier Quick Reference
- Base (145 MB): Fastest transcription (~10x real-time), optimal for MacBook Air with 8 GB RAM.
- Small (480 MB): Balanced accuracy and speed (~5x real-time), recommended for 16 GB machines.
- Medium (1.5 GB): High nuance for complex technical discussions (~2.5x real-time).
- Large-v3 (3.1 GB): Maximum academic precision and multilingual handling on Apple Silicon M-Pro/Max systems.
