On-Device Speech Models

A 1.5B-parameter speech model now runs locally on an iPhone in roughly 2.2GB of memory at up to 1.28 times real-time speed. This supplies the missing device benchmark for a watch line previously supported by offline narration and voice-cloning workflows. With five publications across three source groups and three observed days, edge speech now looks less like a model-compression demo and more like a deployable interface layer for private, low-latency applications.

An open-source tool now performs multilingual subtitle narration, voice cloning, and duration matching fully offline in a lightweight package. This adds a concrete creator workflow to the existing edge-speech hypothesis, but the line still lacks enough distinct products and device benchmarks for publication. Confirm through additional browser or device deployments and recurring production use.

Very small speech models are being positioned for real-time use on CPUs, browsers and single-board devices, while research targets lightweight speech understanding for edge deployment. The early hypothesis is that speech interfaces will become a local capability rather than a cloud-only feature. Confirm with more downloadable models, device benchmarks and concrete workflows.