AI-Powered Noise Awareness in Noise-Cancelling Headphones

In this work, a robust and lightweight deep learning-based system was developed for detecting horn sounds in noisy environments using spectrogram-based feature extraction and convolutional neural networks (CNNs). The system was trained on a carefully curated dataset comprising both horn and non-horn sounds, including various environmental noises such as human speech, birds, etc. Through augmentation techniques like time shifting, the dataset was enhanced to improve model generalization and robustness. A key aspect of the system involved transforming raw audio signals into linear Mel spectrograms, which preserve the temporal and frequency characteristics of sounds in a visually interpretable format. These spectrograms served as the primary input to the CNN model, enabling the network to learn spatial patterns relevant to distinguishing between horn and non-horn audio events. The CNN architecture was carefully designed with multiple convolutional, pooling, and dense layers to progressively extract and process hierarchical features from the input spectrograms. The model demonstrated strong classification performance, effectively capturing subtle differences in frequency distributions and temporal variations between the two classes. To enable real-world deployment, the trained model underwent full integer quantization using a representative dataset. This quantization significantly reduced the memory and computational requirements of the model without substantially impacting accuracy. The quantized model was converted into a MicroTFLite format and integrated into the ESP32-S3 N16R8 microcontroller via the Arduino IDE. Both the model and the input features were embedded into the firmware as C arrays, and real-time inference was carried out directly on the device. The successful execution of the model on the ESP32 platform highlighted the feasibility of deploying deep learning-based audio classification systems on low-power, resource-constrained embedded devices. The system's design and implementation demonstrate a practical and scalable approach to environmental sound detection, especially in applications such as smart noise-cancelling headphones, or urban acoustic monitoring. By enabling real-time detection of safety-critical auditory events like vehicle horns, the solution contributes to enhanced situational awareness in noisy environments. Future work may include expanding the dataset to encompass a broader range of horn types and non-horn noises, improving noise resilience through advanced signal processing or attention-based models, and integrating on-device audio preprocessing to enable end-to-end inference directly from raw microphone input. Additionally, power consumption and inference latency can be further optimized for continuous operation in wearable or battery-powered systems. Overall, this work establishes a strong foundation for deploying intelligent, low-cost, and real-time sound classification systems at the edge.

S. N. C. K. S.Chu, “Environmental Sound Recognition with Time-Frequency Audio Features,” IEEE Transactions on Audio, Sppech, and Language Processing, pp. 1142-1158, 2009. J. L. S. T. Boyang Zhang, “Audio Recognition using Mel Spectrograms and convolutional neural networks,” IEEE, 2010. K. J. Piczak, “"Environmental sound classification with convolutional neural networks,” IEEE, 2015. M. A. S. L. K.-R. M. W. S. Soren Becker, “Interpreting and explaining deep neural networks for classification of audio signals.,” 2018 C. D. S. Q. J. L. a. S. D. W. Dai, “Very deep convolutional neural networks for raw waveforms,” in IEEE International Conference on Acoustics, Speech and Signal Processing, 2017. D. Chenna, “Edge AI: Quantizatuion as the key to on-device smartness,” International Journal of Artificial Intelligence & Applications (IJAIAP) , vol. Volume 2, no. 2, Issue 01, pp. pp. 58-69, 2023. Z. L. R. B. a. D. M. R. Ding, “Quantized deep neural networks for energy efficient hardware-based inference,” IEEE, p. pp. 1–8, 2018. S. K. B. C. M. Z. M. T. A. H. H. A. D. K. B. Jacob, “Quantization and training of neural networks for efficient integer-arithmetic-only inference,” in IEEE Conference on Computer Vision and Pattern Recognition, 2018. S. G. Erez Manor, “Custom Hardware Inference Accelerator for TensorFlow Lite for Microcontrollers,” in IEEE Access, 2022 W.-H. T. Y. F. M. G. Van-Thuan Tran, “End-to-End Train Horn Detection for Railway Transit Safety,” in MDPI, 2022.