Source-linked AI summary
End-to-End Environmental Sound Classification using a 1D Convolutional Neural Network
Sajjad Abdoli, Patrick Cardinal, Alessandro Lameiras Koerich
TL;DR
Environmental sound classifiers often rely on handcrafted or 2D representations, while fixed-input CNNs face variable-duration audio and limited labeled data. This paper develops a compact end-to-end 1D CNN that learns from waveforms, supports variable-length signals with sliding windows, and optionally uses Gammatone initialization. On UrbanSound8k, it achieves 89% mean accuracy and outperforms the compared 2D-representation approaches, while using fewer parameters. The authors note possible complementarity with 2D methods and noisy intermediate-layer filters as areas for further investigation.
Problem
Environmental sound classification commonly uses handcrafted or 2D representations, while CNNs require fixed input lengths and labeled environmental-sound data are scarce.
Method
The paper proposes a compact end-to-end 1D CNN that learns representations directly from audio waveforms, handles variable lengths with sliding-window frames, and can initialize its first layer with Gammatone filters.
Results
89% mean accuracy was achieved on UrbanSound8k, outperforming the compared approaches based on 2D representations.
Takeaways & Limitations
The proposed approach provides state-of-the-art environmental sound classification with fewer parameters and without data augmentation or signal preprocessing.
Takeaways & Limitations
The authors identify possible complementarity with 2D representations and noisy intermediate-layer filters, requiring further investigation.
Abstract
from arXiv · showhide
In this paper, we present an end-to-end approach for environmental sound classification based on a 1D Convolution Neural Network (CNN) that learns a representation directly from the audio signal. Several convolutional layers are used to capture the signal's fine time structure and learn diverse filters that are relevant to the classification task. The proposed approach can deal with audio signals of any length as it splits the signal into overlapped frames using a sliding window. Different architectures considering several input sizes are evaluated, including the initialization of the first convolutional layer with a Gammatone filterbank that models the human auditory filter response in the cochlea. The performance of the proposed end-to-end approach in classifying environmental sounds was assessed on the UrbanSound8k dataset and the experimental results have shown that it achieves 89% of mean accuracy. Therefore, the propose approach outperforms most of the state-of-the-art approaches that use handcrafted features or 2D representations as input. Furthermore, the proposed approach has a small number of parameters compared to other architectures found in the literature, which reduces the amount of data required for training.
1. Introduction
Environmental sound classification has practical applications, but existing approaches often rely on handcrafted or 2D spectrogram representations and substantial training data. The paper proposes a compact end-to-end 1D CNN that learns directly from waveforms and handles variable-length audio.
- Environmental sound classification supports applications including crime detection and urban acoustic monitoring.
- Most existing systems use handcrafted features or mid-level spectro-temporal representations, while CNN approaches commonly use spectrograms as 2D inputs.
- 2D CNNs can require large training datasets to learn kernel parameters without over-fitting, while labeled environmental sound data are scarce.
- The proposed end-to-end 1D CNN learns representations directly from audio waveforms instead of 2D representations.
- On UrbanSound8k, the approach outperformed 2D-representation methods by 11.24% to 27.14% in mean accuracy and required neither data augmentation nor signal preprocessing.The dataset contains 8,732 environmental sounds from 10 classes.
- The architecture uses few parameters, avoids a large training-data requirement, and handles audio signals of any length through sliding-window framing.The first convolutional layer can be initialized with Gammatone filterbanks.
2. Proposed End-to-End Architecture
The proposed architecture is an end-to-end 1D CNN that learns directly from audio waveforms while accommodating variable-length signals through overlapping sliding-window frames. It uses a compact convolutional topology, optionally initialized with Gammatone filters, and aggregates frame-level predictions for classification.
- Variable Audio Length: The approach targets variable-length environmental audio by splitting waveforms into fixed-length, potentially overlapping frames compatible with the CNN input.A sliding window of variable width conditions signals to the input layer; successive frames can overlap to maximize information use.
- Variable Audio Length: A 16 kHz sampling rate is presented as a trade-off between environmental-sound input quality and model computational cost.Sampling rate directly affects input dimensionality and computational cost.
- 1D CNN Topology: The CNN maps raw waveform arrays to class predictions through hierarchical trainable feature extraction, convolution, pooling, and softmax output.Convolution uses one-dimensional kernels and biases, while pooling expands the area covered by subsequent receptive fields; softmax outputs class-membership degrees.
- 1D CNN Topology: The baseline topology uses four convolutional layers, possible max-pooling layers, two fully connected layers, and an output layer.The baseline input has 16,000 dimensions, representing one second sampled at 16 kHz; other input lengths require architectural adaptation, padding, or segmentation.
- 1D CNN Topology: For long recordings, shorter frames are preferred to enlarging the network input, because enlarging it increases parameters and model complexity.The architecture has few parameters, and deeper architectures are avoided because limited training data makes substantial over-fitting impractical.
- Gammatone Filterbanks: The first convolutional layer can be initialized with 64 band-pass Gammatone filters spanning central frequencies from 100 Hz to 8 kHz.This initialization decomposes the waveform into 64 frequency bands and provides a trade-off between handcrafted features and representation learning.
3. Experimental Results
Experiments on UrbanSound8k evaluate architecture depth, input length, learned filters, and three enhancements for the end-to-end 1D CNN. The 16,000-input model provides the strongest accuracy–parameter trade-off, while windowing and Gammatone initialization improve mean accuracy.
- Architecture depth: 69%, 75%, 79% and 80% test accuracy were obtained with one, two, three and four convolutional layers, respectively.Four layers were the upper limit because the minimal feature-map dimension was reached.
- Input lengths and evaluation: The 16,000-input architecture matched the accuracy of the 59,999-input network with almost half its parameters and exceeded the other input-size configurations.The 8,000-input architecture achieved a mean accuracy close to the 16,000-input model with almost three times fewer parameters.
- Input lengths and evaluation: The 16,000-input model had the highest median, smallest interquartile range, and no outlier, making it the best accuracy–parameter trade-off.It had the same mean accuracy as the 50,999-input model with almost half as many parameters.
- Filter analysis: Deeper learned filters had well-defined frequency responses, whereas early filters were noisier; Gammatone first-layer filters improved mean accuracy despite limited intermediate-layer changes.The Fourier analysis covered randomly initialized filters trained for the classification task, with parameters learned directly from data.
- Architecture enhancement: 2% mean-accuracy gains came from the rectangular window, increased overlap from 50% to 75%, and Gammatone initialization, respectively.Gammatone initialization doubled the number of parameters by increasing the first-layer filter count, yet the 16,000-input model remained best.
4. Discussion
On UrbanSound8k, the proposed end-to-end 1D CNN achieves the strongest reported classification performance while using fewer parameters than most alternatives. Its learned filters show frequency-selective behavior, though intermediate-layer filters remain noisy and class-level complementarities with 2D methods may exist.
- 89% mean accuracy was achieved across the 10 UrbanSound8k folds, with a standard deviation of 0.9%.
- The proposed 1D CNN outperformed the other approaches in Table 4 while learning representations directly from audio waveforms.Most other listed approaches use 2D audio representations as input.
- ST and CH were the hardest classes, whereas EN and GU were well separated by the proposed CNN.
- Compared with the SB-CNN, the proposed CNN performed better for AI, DR, EN, SI, CA, and JA, while SB-CNN was slightly better for CH and DO.
- The first-layer learned filters exhibited a logarithmic response resembling band-pass filters and human sound perception.The response was examined using sinusoidal inputs and compared with Gammatone filters.
5. Conclusion
The paper concludes that an end-to-end 1D CNN can learn relevant representations directly from environmental audio and outperform most evaluated 2D CNN architectures with fewer parameters. It also identifies potential complementarity between 1D and 2D representations and noise in intermediate-layer filters as directions for further work.
- The proposed end-to-end 1D CNN learns filters directly from audio waveforms instead of using handcrafted static filterbanks.The architecture contains three to five convolutional layers depending on audio-signal length.
- The 1D architecture achieved mean accuracy 11.24% to 27.14% higher than most evaluated 2D architectures while using fewer parameters.
- The authors identify possible complementarity between learned 1D and 2D filters for some sound classes.They propose investigating whether combining both representations can improve classification.
- Intermediate convolutional-layer filters appeared noisy and lacked dominant frequencies, motivating further investigation.
Availability of Data and Material
The UrbanSound8k dataset was used for training and testing, and the authors state that the proposed CNN source code will be released in the final paper version.
- UrbanSound8k was used to train and test the method.The dataset is available online.
- The proposed end-to-end CNN source code was planned for release in the final version of the paper.
Credit Authorship Contribution Statement
The contribution statement assigns the first author primary responsibility across the research workflow, while the other authors contributed to methodology, validation, supervision, and project-related activities.
- Sajjad Abdoli contributed to conceptualization, methodology, software, validation, analysis, investigation, data curation, writing, and visualization.
- Patrick Cardinal and Alessandro Lameiras Koerich contributed to conceptualization, methodology, validation, analysis, investigation, resources, data curation, writing, supervision, project administration, and funding acquisition.