Fact · Speech recognition
47percent less energy use
Li and colleagues report that their component-level design changes cut the energy use of an on-device streaming speech recognition model by up to 47 percent while keeping accuracy comparable.
Checked against its source on October 17, 2026
In the source's own words
“achieves up to 47% lower energy usage while preserving comparable model accuracy”
Breaking Down Power Barriers in On-Device Streaming ASR: Insights and Solutions, Yang Li et al., arXiv (industry track of NAACL 2025, per the arXiv comments), 2024, Abstract.
Open the original- What it measures
- The paper measures power draw of a streaming model that runs continuously on a device. It shows that how often a component runs, and where its weights sit in memory, both change power use.
- What it does not show
- The 47 percent is a best case against the authors' baselines, on their own model. It is not a battery-life figure for a phone and it does not compare with the energy of sending audio to a server.