PaperlayerView exact revision on arXiv

Source-linked AI summary

End-to-End Environmental Sound Classification using a 1D Convolutional Neural Network

Sajjad Abdoli, Patrick Cardinal, Alessandro Lameiras Koerich

arXiv:1904.08990v1cs.SDcs.LGstat.ML

TL;DR

Environmental audio varies in duration, while CNNs require fixed-length inputs. The paper proposes an end-to-end 1D CNN that learns from audio waveforms, and reports 89% mean accuracy on UrbanSound8k.

  • Problem

    CNNs require fixed-length inputs, but environmental sounds vary in duration and may require continuous prediction.

    Source: §2.1 · p. 5

  • Method

    The paper proposes an end-to-end 1D CNN that learns representations directly from audio waveforms and uses overlapped sliding-window frames to handle variable-length inputs.

    Source: §1 · p. 4, §2.1 · p. 5

  • Results

    89% mean accuracy, with a 0.9% standard deviation across 10 folds, was achieved on UrbanSound8k, outperforming other state-of-the-art approaches based on 2D representations.

    Source: §4 · p. 15, §5 · p. 19

  • Takeaways & Limitations

    The approach has fewer parameters than most other CNN architectures and does not require as much training data as dense 2D CNNs.

    Source: §1 · p. 4, §5 · p. 19

    The authors note that intermediate convolutional-layer filters appear noisy and lack dominant frequencies, requiring further investigation.

    Source: §5 · p. 20

Abstract

from arXiv · show

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

The paper proposes a compact end-to-end 1D CNN that learns directly from audio waveforms for environmental sound classification. On UrbanSound8k, it reports higher mean accuracy than the compared 2D-representation methods and can process audio signals of any length using sliding-window frames.

1. Introduction

  • For environmental sound classification, most methods use handcrafted or spectro-temporal features; 2D CNN capacity depends on abundant training data to avoid overfitting, and labeled-data scarcity is addressed with augmentation.Salamon & Bello (2017) trained a 2D CNN on augmented samples and evaluated it on original samples, reporting 79% classification accuracy on an environmental-sounds dataset.Source: §1 · pp. 2–3
  • On UrbanSound8k, the proposed approach exceeds 2D-representation methods by 11.24% (SB-CNN) to 27.14% (VGG) in mean accuracy.The dataset contains 8,732 environmental sounds from 10 classes.Source: §1 · p. 4
  • The compact end-to-end 1D CNN learns low- and high-level representations directly from waveforms, then classifies them with three fully connected layers.The approach requires neither data augmentation nor signal preprocessing for feature extraction.Source: §1 · p. 4
  • Initialized with Gammatone filterbanks, the 1D CNN has few parameters and is presented as requiring less training data than dense 2D CNNs.The paper contrasts its parameter count with dense 2D CNNs that have millions of trainable parameters.Source: §1 · p. 4
  • Sliding-window framing lets the model handle audio signals of any length by splitting them into short frames compatible with the CNN input layer.Source: §1 · p. 4

2. Proposed End-to-End Architecture

The proposed end-to-end 1D CNN learns a discriminative representation directly from audio and is designed to handle variable-duration environmental sounds. It adapts fixed-size CNN inputs through overlapping frames and can optionally initialize its first layer with Gammatone filters.

2.1. Variable Audio Length

  • Overlapping sliding-window frames let the CNN process variable-length audio while reusing signal segments to increase the number of training samples.Window width depends mainly on sampling rate; the paper notes that 16 kHz may balance input quality and computational cost.Source: §2.1 · pp. 5–6

2.2. 1D CNN Topology

  • Convolutional layers extract hierarchical features from the waveform using local receptive fields; pooling expands the next layers’ coverage, and the final maps are flattened into fully connected layers.The network uses raw waveform input rather than handcrafted features, with softmax outputs indicating class membership.Source: §2.2 · p. 6
  • The compact baseline has four convolutional layers, optional max pooling, two fully connected layers, and an output layer, with input adaptations for other audio lengths.Its baseline input is 16,000 dimensions, representing one second sampled at 16 kHz; longer recordings can instead be split into shorter frames to keep the network compact.Source: §2.2 · pp. 6–7

2.3. Gammatone Filterbanks

  • Optional first-layer initialization uses 64 band-pass Gammatone filters spanning central frequencies from 100 Hz to 8 kHz, decomposing audio into 64 frequency bands.The authors describe this as a trade-off between handcrafted features and representation learning.Source: §2.3 · p. 9

2.4. Aggregation of Audio Frames

  • For audio split into frames, majority vote or the sum rule can combine frame-level CNN predictions into a final classification.The selected class is the one with the maximum aggregated value across the classes.Source: §2.4 · pp. 9–10

3. Experimental Results

Experiments on UrbanSound8k evaluate CNN depth, input length and architecture enhancements. The 16,000-input model offers the strongest reported accuracy–parameter trade-off, and three tested enhancements each improve mean accuracy.

3. Experimental Results

  • UrbanSound8k contains 8,732 clips totaling 7.3 hours, each at most four seconds long; experiments downsampled recordings to 16 kHz.The dataset has ten environmental-sound classes, with class counts ranging from 374 to 1,000 clips.Source: §3 · p. 11

3.1. Fine-Tuning the 1D CNN Architecture

  • Test accuracy rose from 69% with one convolutional layer to 80% with four, the upper limit because the feature map reached its minimum dimension.The exploratory experiment used 16,000-sample segments with 50% overlap; the same procedure informed layer choices for other configurations.Source: §3.1 · p. 11

3.2. Evaluation on Different Audio Lengths

  • The 16,000-input CNN matched the 59,999-input model’s highest mean accuracy; 8,000 inputs came close with nearly three times fewer parameters, while 1,600 inputs scored about 7% lower.The 16,000-input model also had the highest median, smallest interquartile range, and no outliers; it used almost half the parameters of the 50,999-input second-best choice. Audio was segmented into frames of 1,600, 8,000, 16,000, or 32,000 samples, and frame predictions were combined by majority voting or the sum rule.Source: §3.2 · pp. 11–12, Table 2 · p. 12, Figure 5 · p. 13
  • Filters in the first two convolutional layers were noisy and lacked dominant frequencies, while deeper-layer filters had more regular frequency responses.In the enhanced model, CL3 and CL4 retained well-defined responses, while CL2 still lacked dominant frequencies; deeper filters had lower Fourier resolution because they were smaller.Source: §3.2 · pp. 12–13, Figure 6 · p. 14, §3.3 · p. 15

3.3. Architecture Enhancement

  • Rectangular windows, increasing overlap from 50% to 75%, and Gammatone initialization each improved mean accuracy by 2%; Gammatone initialization doubled the parameter count.These enhancements also improved most other architectures, but the 16,000-input CNN remained highest in mean accuracy.Source: §3.3 · pp. 13, 15, Table 3 · p. 15

4. Discussion

On UrbanSound8k, the proposed end-to-end 1D CNN achieved 89% mean accuracy and outperformed the compared approaches; its learned filters showed a logarithmic frequency response similar to human sound perception.

4. Discussion

  • 89% mean accuracy with a 0.9% standard deviation across 10 folds; the proposed 1D CNN outperformed all other approaches in Table 4.Figure 8 also shows the proposed CNN had the highest mean and median accuracy, one of the smallest interquartile ranges, and no outlier. It has fewer parameters than most compared approaches, suggesting it can be trained with fewer samples; unlike most Table 4 approaches, it learns directly from waveforms rather than 2D representations.Source: §4 · p. 15, Table 4 · p. 16, Figure 8 · p. 17
  • The proposed CNN matched SB-CNN on GU and ST, outperformed it on AI, DR, EN, SI, CA, and JA, and was slightly worse on CH and DO.SB-CNN with data augmentation (Salamon & Bello, 2017) used spectrograms and needed 20 times more samples to achieve 79% accuracy; the authors suggest fusing the classifiers because their per-class accuracies differ, including on AI and JA.Source: §4 · p. 18, Figure 10 · p. 18
  • ST and CH were the hardest classes for the proposed CNN, while EN and GU were well separated.The confusion matrix reports the number of correctly classified samples along its diagonal.Source: §4 · pp. 15, 18, Figure 9 · p. 17

4.1. Filter Response

  • The learned first-layer filters had a logarithmic frequency response, resembling both band-pass filters and human sound perception.The response analysis used 64 first-layer kernels rather than the experimental 16, a configuration that slightly decreased classification accuracy; sinusoidal probes ranged from 1 Hz to 8 kHz in 100 Hz steps, and Figure 11(c) shows 64 Gammatone band-pass filters.Source: §4.1 · pp. 18–19, Figure 11 · p. 19

5. Conclusion

The paper proposes an end-to-end 1D CNN that learns filters from audio waveforms and outperforms state-of-the-art approaches based on 2D representations. It uses fewer parameters than most other environmental-sound CNN architectures while achieving higher mean accuracy.

5. Conclusion

  • Mean accuracy was between 11.24% and 27.14% higher than that of the 2D architectures, while the proposed 1D CNN used fewer parameters.The approach also outperformed other state-of-the-art methods based on 2D representations and 2D CNNs.Source: §5 · p. 19
  • The 1D CNN learns filters directly from waveforms rather than using handcrafted static filterbanks, and its architecture uses three to five convolutional layers depending on audio length.The approach was evaluated on 8,732 audio samples and learned several relevant filter representations.Source: §5 · p. 19
  • Combining learned 1D filters with filters learned from spectrograms may improve overall performance, a possibility the authors identify for future investigation.The paper notes complementarity for at least some classes, as highlighted in Figure 10.Source: §5 · p. 20
  • Intermediate-layer filters lack dominant frequencies and appear noisy, so further investigation is needed to address this issue and potentially improve performance.Source: §5 · p. 20

Availability of Data and Material

The paper uses UrbanSound8k for training and testing, and says its source code will be made available in the final version.

Availability of Data and Material

  • UrbanSound8k is used for training and testing, and the proposed CNN source code will be made available in the final version.The dataset is available at https://urbansounddataset.weebly.com/urbansound8k.html.Source: Availability of Data and Material · p. 20

Related papers

Loading 1904.08990v1…