Fixed-Point Implementation of a Speech Enhancement Neural Network
Zhu, Yue (2026)
2026
Master's Programme in Computing Sciences and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2026-06-24
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202606247915
https://urn.fi/URN:NBN:fi:tuni-202606247915
Tiivistelmä
Deep learning has achieved significant success in speech enhancement. However, deploying deep neural networks on resource-constrained embedded devices remains challenging due to their high memory consumption and intensive computational requirements. Furthermore, certain hardware platforms strictly support integer arithmetic.
To address these challenges, this study presents a manual conversion of a Python- based speech enhancement model into a C implementation for hardware deployment, applying fixed-point quantization to achieve efficient model compression. To mitigate the precision degradation typically caused by standard static quantization, a dynamic fixed-point quantization strategy is introduced to adaptively preserve numerical accuracy. The performance of the quantized C models are evaluated using a variety of evaluation metrics.
Experimental results demonstrate that the proposed fixed-point C model successfully maintains high numerical fidelity compared to the floating-point baseline, while the dynamic fixed-point strategy effectively achieves higher precision than the standard static fixed-point approach.
To address these challenges, this study presents a manual conversion of a Python- based speech enhancement model into a C implementation for hardware deployment, applying fixed-point quantization to achieve efficient model compression. To mitigate the precision degradation typically caused by standard static quantization, a dynamic fixed-point quantization strategy is introduced to adaptively preserve numerical accuracy. The performance of the quantized C models are evaluated using a variety of evaluation metrics.
Experimental results demonstrate that the proposed fixed-point C model successfully maintains high numerical fidelity compared to the floating-point baseline, while the dynamic fixed-point strategy effectively achieves higher precision than the standard static fixed-point approach.