Whisper benchmarks on Apple silicon
On-device speech transcription requires high throughput without monopolizing system resources. Because Apple silicon integrates CPU, GPU, and Neural Engine on a unified memory bus, Whisper runs at exceptional speeds with low thermal footprint.
Measured benchmarks across Apple silicon chips
The table below reflects measured batch transcription times for a 60-minute clean English audio recording across Mac configurations:
| Apple Silicon Chip | Whisper Model | RAM Used | Time to Complete | Real-Time Factor (RTF) |
|---|---|---|---|---|
| Apple M4 | Base (148 MB) | ~320 MB | 1 min 42 sec | 0.028x |
| Apple M4 | Small (488 MB) | ~850 MB | 3 min 15 sec | 0.054x |
| Apple M4 | Large v3 Turbo (1.6 GB) | ~2.1 GB | 5 min 50 sec | 0.097x |
| Apple M3 Pro | Small (488 MB) | ~850 MB | 3 min 45 sec | 0.062x |
| Apple M3 Pro | Large v3 Turbo | ~2.1 GB | 6 min 40 sec | 0.111x |
| Apple M2 Air | Base (148 MB) | ~320 MB | 2 min 10 sec | 0.036x |
| Apple M2 Air | Small (488 MB) | ~850 MB | 4 min 20 sec | 0.072x |
| Apple M1 | Base (148 MB) | ~320 MB | 2 min 40 sec | 0.044x |
| Apple M1 | Small (488 MB) | ~850 MB | 5 min 10 sec | 0.086x |
Real-Time Factor (RTF): Processing time divided by audio duration. An RTF of 0.05x means 1 hour of audio takes 3 minutes to transcribe.
Word Error Rate (WER) comparison
| Model | Model Size | Typical English WER | Ideal Use Case |
|---|---|---|---|
| Base | 148 MB | ~8.5% | Fast 1-on-1 syncs, daily standups, quiet rooms |
| Small | 488 MB | ~5.8% | General meetings, multi-speaker calls, light accents |
| Large v3 Turbo | 1.6 GB | ~3.9% | Complex technical architecture reviews, heavy accents, legal/medical terminology |
Memory allocation recommendations
- 8 GB Unified Memory Macs: We recommend sticking to Base or Small. Running Large v3 Turbo simultaneously with local Ollama on 8 GB can cause memory pressure.
- 16 GB+ Unified Memory Macs: Easily handles Large v3 Turbo alongside 8B local LLMs (such as Llama 3) without system paging.
Related guides
- Model sizing in Transcription models.
- Audio batch processing in Batch audio and video transcription.
- RAM management in Memory consumption and optimization.
Questions
How fast does Whisper run on Apple silicon?
On an M4 Mac, Whisper Base transcribes 60 minutes of audio in under 2 minutes. Large v3 Turbo transcribes 60 minutes in about 6 minutes.
What is Word Error Rate (WER)?
WER measures the percentage of incorrectly transcribed words. Lower is better. Large v3 Turbo achieves sub-5% WER on clean speech.
Does Whisper use the Apple Neural Engine?
Yes. Oyma utilizes Metal and CoreML optimizations to execute Whisper layers across the Apple Neural Engine and GPU.
