Resource‐Constrained Keyword Spotting With Quantized CNNs: Implementation Strategies and Performance Trade‐Offs.
Saved in:
| Title: | Resource‐Constrained Keyword Spotting With Quantized CNNs: Implementation Strategies and Performance Trade‐Offs. |
|---|---|
| Authors: | Medur, Patrik1 (AUTHOR), Lubbers, Mart2 (AUTHOR), Mauša, Goran1 (AUTHOR) goran.mausa@uniri.hr |
| Source: | International Journal of Intelligent Systems. 8/31/2026, Vol. 2026, p1-10. 10p. |
| Subjects: | Convolutional neural networks, Embedded computer systems, Internet of things, Machine learning, Speech perception |
| Abstract: | Resource‐constrained embedded devices require specialized optimization strategies to achieve reliable real‐time keyword spotting performance. This work presents a comprehensive implementation framework for deploying quantized convolutional neural networks on the ESP32‐S3‐EYE development board, achieving practical capabilities within severe memory and computational limitations. The implementation utilizes Google's Speech Commands Dataset V0.02 with MFCC preprocessing, following the standard 12‐class evaluation convention, 10 target keywords together with 2 auxiliary classes to ensure comparability with established KWS benchmarks. The deployed model achieves 89.58% accuracy in laboratory conditions with a quantized TensorFlow Lite (LiteRT) memory of 0.68 MB, enabling deployment on edge devices with limited memory resources. A dual‐core system architecture implements FreeRTOS‐based task distribution, dedicating one core to continuous audio capture, while the second handles feature extraction and inference operations. Voice activity detection with adaptive noise floor monitoring triggers processing when speech exceeds background noise levels. Comprehensive validation reveals substantial performance differences between controlled and real‐world conditions. Testing with 13 non‐native English speakers demonstrates 81.82% recognition accuracy. Synthetic speech evaluation using 10 AI voices achieves 90.06% accuracy, confirming that acoustic consistency significantly impacts recognition performance. The tiny machine learning framework enables practical keyword spotting on resource‐constrained embedded systems for IoT applications. [ABSTRACT FROM AUTHOR] |
| Copyright of International Journal of Intelligent Systems is the property of Wiley-Blackwell and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) | |
| Database: | Engineering Source |
|
Full text is not displayed to guests.
Login for full access.
|
|
| Abstract: | Resource‐constrained embedded devices require specialized optimization strategies to achieve reliable real‐time keyword spotting performance. This work presents a comprehensive implementation framework for deploying quantized convolutional neural networks on the ESP32‐S3‐EYE development board, achieving practical capabilities within severe memory and computational limitations. The implementation utilizes Google's Speech Commands Dataset V0.02 with MFCC preprocessing, following the standard 12‐class evaluation convention, 10 target keywords together with 2 auxiliary classes to ensure comparability with established KWS benchmarks. The deployed model achieves 89.58% accuracy in laboratory conditions with a quantized TensorFlow Lite (LiteRT) memory of 0.68 MB, enabling deployment on edge devices with limited memory resources. A dual‐core system architecture implements FreeRTOS‐based task distribution, dedicating one core to continuous audio capture, while the second handles feature extraction and inference operations. Voice activity detection with adaptive noise floor monitoring triggers processing when speech exceeds background noise levels. Comprehensive validation reveals substantial performance differences between controlled and real‐world conditions. Testing with 13 non‐native English speakers demonstrates 81.82% recognition accuracy. Synthetic speech evaluation using 10 AI voices achieves 90.06% accuracy, confirming that acoustic consistency significantly impacts recognition performance. The tiny machine learning framework enables practical keyword spotting on resource‐constrained embedded systems for IoT applications. [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 08848173 |
| DOI: | 10.1155/int/2317323 |
