Best local speech-to-text models for CPU-only PCs
On CPU-only hardware, model size and runtime matter as much as nominal model quality. The goal is to avoid choosing a model that is technically local but painfully slow.
Best multilingual starting point: Whisper small or base
Whisper base and small are mature and broadly supported by local runtimes such as whisper.cpp and faster-whisper. Start with base when responsiveness matters, or small when you can spend more compute for better transcription.
Best tiny English option: Moonshine Tiny
Moonshine Tiny is an English-focused 26M-class model designed for constrained hardware. It is a strong candidate for voice commands and lightweight local applications when English is enough.
When not to force CPU-only
Large multilingual models can still run through quantized runtimes, but “it runs” is not the same as “it is a good user experience.” For long daily recordings, a modest GPU or Apple Silicon can change the practical answer.
Checklist
- Test real-time factor on your own CPU.
- Use representative noisy audio, not one clean demo clip.
- Measure peak RAM as well as average speed.
- Keep a smaller fallback model for live interactions.
Reviewed against primary sources on September 13, 2026. Model behavior and licenses can change; re-check the linked model source before production use.