A dynamic template adaptation approach for noise-robust sound classification and distance determination in single-channel audio

In this study, we propose a dynamic template adaptation approach for noise-robust sound classification and distance estimation in single-channel audio environments. Traditional cross-correlation methods rely on fixed sound templates that limit their performance under dynamic and noisy conditions. Our method integrates a low-pass filter for noise reduction and uses an online support vector machine (SVM) to dynamically update the sound templates based on real-time audio inputs. This hybrid approach enables continuous refinement of the templates and improves both the accuracy of sound classification and the ability to determine the relative distance between sound sources by estimating time delays. The robustness and adaptability of the algorithm make it suitable for real-world applications such as environmental monitoring, speaker recognition, and sound event localization. We demonstrate the effectiveness of the proposed method in various noisy and overlapping sound scenarios and compare it with traditional approaches such as ICA, NMF, and TFM. The results show that dynamic template adaptation and incremental learning significantly improve the classification accuracy and distance detection in changing environments. These findings demonstrate that the proposed method not only enhances real-time sound classification and distance determination but also holds potential for applications in autonomous vehicles, urban noise monitoring, and smart home systems, where robust audio processing in dynamic environments is critical.