---
id: 20261031-T0-17
title: "通过循环实现深度：迈向超高效端侧ASR"
title_en: "Looping structure replaces stacked layers to enable ultra-efficient on-device ASR"
url: https://ai.daily.yangsir.net/daily/20261031-T0-17
issue_date: 2026-10-31
publish_date: 2026-10-30T04:00:00.000Z
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2610.37833
---

# 通过循环实现深度：迈向超高效端侧ASR

arXiv:2610.37833v1 公告类型：新论文 摘要：现代自动语音识别（ASR）系统通过扩展模型深度和容量实现了显著准确率，但代价是内存、能耗和计算需求增加。在ASR往往最需要的边缘设备上，例如手表等可穿戴设备，这种资源开销成为关键瓶颈。该研究提出以循环结构替代单纯堆叠层数来获得模型深度，从而在保持识别精度的同时大幅降低参数量与推理成本，目标是在资源受限的端侧设备上实现超高效ASR部署。

## English Version

**Looping structure replaces stacked layers to enable ultra-efficient on-device ASR**

arXiv:2610.37833v1 (new paper) proposes achieving model depth through recurrent structure rather than simply stacking layers, aiming to enable ultra-efficient on-device ASR. Modern ASR systems gain accuracy by scaling model depth and capacity, but at the cost of higher memory, energy, and compute demands. On edge devices such as wearables and watches, where ASR is often most needed, this resource overhead becomes a critical bottleneck. The research seeks to preserve recognition accuracy while substantially reducing parameter count and inference cost, targeting deployment of ultra-efficient ASR on resource-constrained edge devices.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2610.37833)

**详情页**：https://ai.daily.yangsir.net/daily/20261031-T0-17

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*