Year: 2026
Type: article
Abstract: The following paper refers to a lightweight end-to-end speech-to-text system targeted at running on low-power edge hardware systems. The architecture suggested will utilise both an implementation as a small convolutionalrecurrent neural network and Connectionist Temporal Classification (CTC) decoding to facilitate the achievement of real-time automatic speech recognition within a set of stringent memory and energy limits. In order to reduce computational burden and still maintain the accuracy of recognition, the network is compressed by applying structured pruning and post-training quantization (8bits) to the network. After quantizing its model, 1.2 million parameters are contained in the optimised model which occupy 4.8 MB of flash memory. Video-based experiments on an ARM-based edge device with low power reveal real time inference at less than 120ms per utterance. With compression, power consumption is reduced by 38 percent with comparison to the implementations under baseline where there is no compression. A baseline speech evaluation distance results in WordErrorRate (WER) of 9.8% which implies that state of the art recognition systems can operate well in limited hardware settings. The findings prove that appropriate design of lightweight neural networks, as well as hardware-aware optimization strategy, can produce accurate speech recognitions under low-energy consumption in embedded devices. This architecture is scaled and can be used in smart home systems, IoT devices and portable assistive devices that need to create reliable speech processing on the device.
Author: Chuong Van, MH Trinh, T Shimada
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7220094039