Preprint
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
Read the originalRead on October 17, 2026
- What it is
- An industry study of how the power draw of an on-device streaming speech recognition model depends on its weights and how often each component runs. The authors compress model components by their energy sensitivity.
- How much weight it carries
- A preprint from a large technology company that lists a NAACL 2025 industry track venue in its comments. The 47 percent figure compares against the authors' own baselines on their own hardware and model, so it is not a battery-life measurement on a phone.
Facts taken from this source
- 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.
“achieves up to 47% lower energy usage while preserving comparable model accuracy” (Abstract)