What question did this study set out to answer?

The aim is to classify automotive horns based on sound quality using machine learning and deep learning, reducing reliance on time-consuming jury tests.

February 26, 2026

Classification of automotive horns using multi-feature combination with machine learning and deep learning methods

Key Points

The aim is to classify automotive horns based on sound quality using machine learning and deep learning, reducing reliance on time-consuming jury tests.
Recorded 22 horn sound samples using a Binaural Head Measurement System (BHMS).
Applied audio feature extraction techniques like psychoacoustics, mel-frequency cepstral coefficients, and time-frequency features.
Utilized design of experiments (DOE) to determine key features for classification.
Employed noise augmentation, stretching, and shifting on audio samples to enhance data diversity.
Implemented SMOTE to address sample balance issues in classification datasets.
Random Forest and Convolutional Neural Network emerged as the highest-performing models for horn classification.
CNN model exhibited high training time compared to other methods, highlighting a trade-off between performance and efficiency.
Using the developed models can significantly reduce the reliance on traditional jury tests during horn design.

Abstract

Honking contributes significantly to traffic noise but automotive horns are selected mainly based on sound pressure level (SPL) without considering sound quality (SQ). Horns with the same SPL may be perceived differently because of variation in sound quality parameters. Traditional SQ study relies on jury tests which are time and resource intensive. Present study proposes machine learning (ML) and deep learning (DL) models for horn classification, minimizing the need for jury tests. Previous horn sound classifications were performed using audio features like, psychoacoustics, time, frequency attributes separately measured using single microphone data. As human hearing is a complex phenomenon, consideration of individual features may not capture perception of sound quality. This work explores different feature combinations with ML/DL models for binary or multi-class horn classification. Twenty-two horn sound samples are recorded using Binaural Head Measurement System (BHMS) and then split into segments. After signal augmentation, samples are labeled as ‘pleasant’, ‘unpleasant’ under binary class using subjective rating data of jury tests. Then, audio features like psychoacoustics, time-frequency features, mel spectrogram (MS), mel Frequency cepstral coefficients (MFCCs), and chromagram are calculated. These features are used to obtain a multi-class horn classification dataset using unsupervised clustering. Design of experiments (DOE) strategy is implemented to find the key features. After hyperparameter tuning, effectiveness of all models is evaluated through comparison of Accuracy, F1-score, Precision, Recall, and training time. Significant attention is given for audio augmentation by injecting noise, stretching, pitching, and shifting to original sound samples. Similarly, Synthetic Minority Over-sampling Technique (SMOTE) is applied to address imbalance in sample numbers. Present study reveals that Random Forest (RF) and Convolutional Neural Network (CNN) are the best performing models. However, the CNN model requires high training time. Developed models can be used to classify horn samples at design stage reducing the need for jury tests.

Mark Helpful

Bookmark

Relay