Fitting a 600M-Parameter ASR Model on Smartphone — Memory, Speed, and Hallucinations

Learn how to run a 600M-parameter ASR model on a smartphone, addressing memory limits, speed, and preventing model hallucinations for a seamless offline experience.

Overview

Pico Memo is an offline voice memo app that transcribes speech entirely on-device using a 600M-parameter AI model, with no server, no account, and no network call ever made. It runs on any iPhone with iOS 13+.

For the live demo I’ll show: the app transcribing speech in real time on a physical iPhone, and walk through how the model stays alive and responsive under the tight memory constraints of a mobile device.

Tech stack