Source-linked AI summary
AudioMNIST: Exploring Explainable Artificial Intelligence for Audio Analysis on a Simple Benchmark
Sören Becker, Johanna Vielhaben, Marcel Ackermann, Klaus-Robert Müller, Sebastian Lapuschkin, Wojciech Samek
TL;DR
Deep audio classifiers remain difficult to interpret, motivating better evidence about which features drive their predictions. The paper introduces AudioMNIST, applies LRP to waveform and spectrogram models, and tests relevance-based explanations through input manipulations and a human study. It reports that audible explanations are more interpretable than visual explanations, while acknowledging scope and representation limits.
Problem
Deep neural networks obscure how inputs produce predictions, and audio classification has fewer interpretation studies than image classification.
Method
The paper introduces AudioMNIST, applies Layer-wise Relevance Propagation to waveform- and spectrogram-based classifiers, and evaluates visual and audible explanations.
Results
Audible explanations show superior interpretability over visual explanations in a human user study, while LRP-based feature manipulations can target classifier-relevant features.
Takeaways & Limitations
LRP provides a way to derive audio-classification strategies for model developers, and audible explanations offer an alternative to visual relevance displays for human users.
Takeaways & Limitations
Audible explanations are currently restricted to positive or negative relevance and the time domain because spectrogram phase information is excluded.
Abstract
from arXiv · showhide
Explainable Artificial Intelligence (XAI) is targeted at understanding how models perform feature selection and derive their classification decisions. This paper explores post-hoc explanations for deep neural networks in the audio domain. Notably, we present a novel Open Source audio dataset consisting of 30,000 audio samples of English spoken digits which we use for classification tasks on spoken digits and speakers' biological sex. We use the popular XAI technique Layer-wise Relevance Propagation (LRP) to identify relevant features for two neural network architectures that process either waveform or spectrogram representations of the data. Based on the relevance scores obtained from LRP, hypotheses about the neural networks' feature selection are derived and subsequently tested through systematic manipulations of the input data. Further, we take a step beyond visual explanations and introduce audible heatmaps. We demonstrate the superior interpretability of audible explanations over visual ones in a human user study.
1. Introduction
The paper addresses the challenge of interpreting deep neural networks in audio by introducing AudioMNIST, a public spoken-digit benchmark, and applying post-hoc explanations to waveform and spectrogram models. It also introduces audible heatmaps and evaluates them against visual explanations in a human study.
- Motivation: Deep neural networks are difficult to interpret because their nonlinear hierarchical architectures obscure connections between inputs and outputs.This opacity creates challenges where prediction understanding and verification are essential.
- Dataset and benchmark: AudioMNIST is a publicly available spoken-digit dataset designed as a basic benchmark for audio classification models and XAI algorithms.The dataset supports spoken-digit recognition and speaker-sex classification tasks.
- Approach: The study trains waveform- and spectrogram-based neural networks and uses Layer-wise Relevance Propagation to identify influential input features.LRP is applied to inspect which timepoints or time-frequency components contribute to predictions.
- Explanation formats: The paper extends visual relevance heatmaps with audible heatmaps and compares both explanation formats in a human interpretability study.The study evaluates which format better communicates model reasoning to users.
2. Explainable AI in the audio domain
The paper represents audio either as raw waveforms or spectrograms and applies Layer-wise Relevance Propagation to explain classifier decisions. It then converts relevance scores into visual or audible explanations, while noting limitations in representing spectrogram relevance audibly.
- Audio representations: Raw waveforms encode amplitude over time, whereas spectrograms represent audio in the time-frequency domain.The waveform is one-dimensional; the spectrogram contains complex-valued frequency and time components, with amplitude commonly used for classifier training.
- Audio representations: The Short Time Discrete Fourier Transform constructs spectrograms from overlapping windowed portions of the waveform.Its window function has a defined length and hop size, producing frequency-bin and time-bin components.
- Layer-wise Relevance Propagation: LRP decomposes a network output into input relevance values by propagating contributions backward through the model.Each relevance value represents an input or hidden variable’s contribution to the final prediction, using layer-specific redistribution rules.
- Explanation formats: LRP produces relevance scores for waveform timepoints or spectrogram time-frequency components, which can be communicated visually or audibly.Visual explanations overlay relevance heatmaps on the input, while audible explanations provide an alternative audio-based format.
- Explanation formats: Audible explanation methods can avoid some segmentation and source-separation requirements associated with AudioLIME.AudioLIME quantifies relevance for separated source segments and plays the most relevant segments.
- Explanation formats: Audible explanations currently present either positive or negative relevance and are limited to the time domain because spectrogram phase information is excluded.Without phase information, directly reconstructing waveforms from spectrogram representations is difficult.
3. Results
The paper benchmarks waveform and spectrogram CNNs on AudioMNIST, then uses LRP and input manipulations to examine their feature selection. Visual and perturbation analyses identify task-dependent and high-amplitude features, while a user study finds audible explanations more interpretable than visual ones.
- AudioMNIST dataset and models: AlexNet consistently outperforms AudioNet, while both models achieve test performance well above chance on digit and sex classification.The comparison is qualified because training parameters differ and model performance is not the paper’s primary goal.
- Visual explanations reveal classifier strategies: LRP relevance maps show task-dependent feature selection: identical spectrograms receive different relevance distributions for digit and sex classification.For sex classification, relevance concentrates in lower-frequency bands, whereas digit explanations do not yield deeper classification-strategy insights.
- Visual explanations reveal classifier strategies: AudioNet relevance maps suggest that high-magnitude waveform samples are particularly relevant to the network’s classification decisions.This observation motivates systematic relevance-guided input perturbations.
- Relevance-guided sample manipulation: At 1% signal manipulation, digit accuracy falls from 92.53% to 92% with random selection, 85% with amplitude-based selection, and 77% with LRP-based selection.For sex classification, random manipulation leaves accuracy stable until 60% of the signal is zeroed; relevance- and amplitude-based perturbations have stronger effects.
- Audible explanations and human interpretability: In the user study, audible explanations outperform visual explanations on informedness and markedness for incorrectly classified samples, although values of 0.12 and 0.1 leave room for improvement.For correctly classified samples, the signal-only baseline can outperform the audible explanation because model and user classification strategies may differ.
- Audible explanations and human interpretability: The paper concludes that explanation presentation matters beyond raw relevance computation and that audible explanations are more interpretable for human users in audio applications.The authors suggest concept-based audible explanations as a possible route to further improvement.
4. Conclusion
The paper contributes an open-source spoken-digit dataset and demonstrates LRP-based explanations for audio classifiers. A human study shows that audible explanations are more interpretable than visual explanations for individual audio predictions.
- The paper provides an Open Source dataset of spoken digits in English for research on audio classification and XAI.
- LRP visual explanations support high-level classification strategies from the perspective of model developers.
- The paper introduces audible explanations that follow the established explanation framework used in computer vision.
- A user study shows that audible explanations have superior interpretability to visual explanations for individual audio signals.
- Future work targets more complex audio datasets and improved audible explanations using concept-based XAI methods.
Appendix A. Model details
The appendix details the AudioNet architecture, speaker-disjoint dataset splits, preprocessing, and training protocols for digit and sex classification. It also records the AlexNet training setup used for both tasks.
- Model architecture: AudioNet uses nine weight layers with convolutional, max-pooling, and fully connected layers ending in either 10 digit or 2 sex outputs.
- Dataset splits: Digit classification uses five speaker-disjoint splits, with three for training and separate validation and test splits in five-fold cross-validation.
- Dataset splits: Sex classification uses 24 speakers divided into four disjoint splits, with two training splits and separate validation and test splits.
- Model training: AlexNet training uses stochastic gradient descent with batch size 100, momentum 0.9, learning-rate decay, and gradient clipping at magnitude 5.
- Model training: AudioNet digit classification runs for 50,000 optimization steps, whereas sex classification runs for 10,000 steps with task-specific learning-rate schedules.