Hearing-impaired individuals often face difficulty in recognizing important environmental sounds such as

sirens, alarms, and warning signals, which can affect their safety and situational awareness in daily life. This

paper presents Resonix, an AI-based wearable assistive system designed to identify critical environmental

sounds and provide real-time haptic alerts through vibration patterns. Environmental sound datasets including

ESC-50 and UrbanSound8K were utilized for model development. Five relevant sound classes — siren, dog

bark, alarm, baby cry, and glass break — were selected for classification. Audio preprocessing techniques such

hearing-impaired individuals through a simple and low-cost wearable solution.

Keywords: Convolutional Neural Network (CNN), Wearable Assistive System, Haptic Alert System, ESP32,

Machine Learning, Audio Classification

Hearing-impaired individuals often face challenges in recognizing important environmental sounds such as

alarms, sirens, and warning signals, which can affect their safety and daily activities. Traditional hearing aids

mainly amplify surrounding audio but may not effectively identify critical sound events in noisy environments

[1]. Environmental Sound Classification (ESC) has emerged as an important research area in assistive

technology and audio intelligence applications [2]. Deep learning approaches, especially Convolutional Neural

Networks (CNNs), have shown effective performance in recognizing environmental sounds from audio datasets

sound classification [2], [3]. CNNs automatically learn discriminative temporal and spectral features from audio

signals [3], [4], making them suitable for real-time assistive applications. Motivated by these advancements,

this work explores the use of CNN-based sound recognition integrated with wearable haptic feedback

technology for hearing-impaired users, similar to recent sound accessibility and assistive technology research

[8], [12], [13].

Environmental sound datasets namely ESC-50 and UrbanSound8K were used for model development and

evaluation [1]. Five environmental sound classes — siren, dog bark, alarm, baby cry, and glass break — were

selected for classification. The audio samples were preprocessed using normalization and noise reduction

Audio signals were normalized to maintain consistent amplitude levels and noise reduction techniques

were applied to minimize background interference. The audio recordings were resampled and segmented

into fixed-duration clips before feature extraction.

Mel-Frequency Cepstral Coefficients (MFCCs) and Mel Spectrograms were extracted from the audio

signals. These features effectively represent temporal and spectral characteristics of environmental sounds

and were used as inputs to the classification models.

The proposed CNN architecture consists of two convolutional layers containing 32 and 64 filters respectively

with ReLU activation. Each convolution layer is followed by max-pooling. A fully connected dense layer of

Network (CNN) were implemented for environmental sound classification [2]. The dataset was divided

into training and testing sets for performance evaluation. Accuracy, precision, recall, and F1-score were

used to compare the performance of different models [5].The comparative analysis indicates that the CNN

model achieved superior performance across all evaluation metrics. The trained CNN model was

integrated with an ESP32-based wearable setup containing three microphone modules, two vibration

motors, and an emergency push button. When a sound was detected and classified, the ESP32 generated

vibration alerts to notify hearing-impaired users in real time [12].

The working of the proposed Resonix system begins with the three microphone modules capturing surrounding

environmental sounds from different directions in real time. The acquired audio signals are preprocessed and

The performance of different machine learning and deep learning models was evaluated using accuracy,

precision, recall, and F1-score metrics for environmental sound classification. The comparative analysis was

carried out using KNN, SVM, Random Forest, and the proposed CNN model. Among all the evaluated models,

the proposed CNN model achieved the highest classification accuracy for environmental sound recognition with

an overall accuracy of 93.0%. The improved performance of the CNN model was mainly due to its capability

to automatically learn discriminative and high-level audio features from environmental sound samples without

reported improved performance due to automatic feature learning and deep hierarchical representations [8], [10].

The proposed model achieved a precision of 92.7%, recall of 92.4%, and F1-score of 92.5%, showing balanced

and reliable classification performance across all selected sound classes.

The normalized confusion matrices further illustrate the classification performance of the evaluated models. The

misclassification among environmental sound categories. Random Forest demonstrated comparatively better

class-wise prediction performance with reduced classification errors. In contrast, the proposed CNN model

showed strong diagonal dominance in the normalized confusion matrix, indicating accurate class-wise prediction

with minimal misclassification. Slight confusion was observed between alarm and baby cry sounds due to

similarities in audio frequency characteristics and temporal sound patterns [11].

The trained CNN model was integrated with the ESP32-based wearable setup for real-time environmental sound

The SVM model improved the classification performance and achieved an accuracy of 84.2% by providing better

decision boundary separation among environmental sound categories. The Random Forest model further

enhanced the prediction capability and achieved an accuracy of 88.5% with improved stability and reduced

classification errors across most sound classes. Among all the evaluated models, the proposed CNN model

demonstrated the best overall performance with an accuracy of 93.0%, precision of 92.7%, recall of 92.4%, and

F1-score of 92.5%. The normalized confusion matrices showed strong diagonal dominance for the CNN model,

indicating accurate class-wise prediction with minimal misclassification. The CNN model effectively learned

discriminative audio features from environmental sound samples, resulting in superior recognition performance

compared to traditional machine learning models. The obtained results demonstrate that deep learning-based

environmental sound classification can provide reliable performance for real-time assistive wearable

impaired users to receive alerts without relying on audio signals. The overall hardware setup is lightweight,

portable, and suitable for wearable implementation

4. Y. Su et al., “Environment Sound Classification Using a Two-Stream CNN,” Sensors, 2019.

5. A. Bansal et al., “Environmental Sound Classification: A descriptive review,” Intelligent Systems with

Applications, 2022.

6. Z. Mushtaq et al., “Spectral images based environmental sound classification using CNN,” Applied

Acoustics, 2021.

7. J. Guo et al., “A Deep Attention Model for Environmental Sound Classification,” Applied Sciences,

2022.

8. R. Akter et al., “A hybrid CNN-LSTM model for environmental sound classification,” Digital Signal

Processing, 2025.

9. S. Abdoli et al., “End-to-End Environmental Sound Classification using 1D CNN,” arXiv, 2019.

13. D. Jain et al., “Deep Learning for Sound Accessibility on Smartwatches,” Communications of the

ACM, 2022.

14. A. Ashurov et al., “Environmental Sound Classification Based on Transfer Learning,” Electronics,

2022.

15. Z. Zhang et al., “Deep CNN with Mixup for Environmental Sound Classification,” arXiv, 2018.