Pocket Proof

Inspiration

We wanted private, offline speech-to-text to be practical on mobile Arm devices—not dependent on cloud uploads.

What it does

Pocket Proof transcribes audio locally and demonstrates how quantization makes Whisper faster, smaller, and more memory-efficient.

How we built it

We used Whisper, native Arm64 inference, browser-based local transcription, and a reproducible benchmark comparing FP16 with Q4_0.

Challenges we ran into

Accurately measuring performance, separating browser and native results, and avoiding unsupported optimization claims were the biggest challenges.

Accomplishments that we're proud of

On an Apple M5, Q4_0 achieved 1.55× faster transcription, a 70% smaller model, and 44% lower peak memory, with transparent quality reporting.

What we learned

Optimization is about more than speed. Model size, memory use, accuracy, reproducibility, and honest reporting all matter.

What's next for Pocket Proof

We plan to support more Arm devices, longer recordings, additional models, and broader real-world audio testing.

Built With

• arm