The correlation between the sharpness of loss minima and generalisation in the context of deep neural networks has been subject to discussion for a long time. Whilst mostly investigated in the context of selected benchmark data sets in the area of computer vision, we explore this aspect for the acoustic scene classification task of the DCASE2020 challenge data. Our analysis is based on two-dimensional filter-normalised visualisations and a derived sharpness measure. Our exploratory analysis shows that sharper minima tend to show better generalisation than flat minima -even more so for out-of-domain data, recorded from previously unseen devices-, thus adding to the dispute about better generalisation capabilities of flat minima. We further find that, in particular, the choice of optimisers is a main driver of the sharpness of minima and we discuss resulting limitations with respect to comparability. Our code, trained model states and loss landscape visualisations are publicly available.
Varying conditions between the data seen at training and at application time remain a major challenge for machine learning. We study this problem in the context of Acoustic Scene Classification (ASC) with mismatching recording devices. Previous works successfully employed frequency-wise normalization of inputs and hidden layer activations in convolutional neural networks to reduce the recording device discrepancy. The main objective of this work was to adopt frequency-wise normalization for Audio Spectrogram Transformers (ASTs), which have recently become the dominant model architecture in ASC. To this end, we first investigate how recording device characteristics are encoded in the hidden layer activations of ASTs. We find that recording device information is initially encoded in the frequency dimension; however, after the first self-attention block, it is largely transformed into the token dimension. Based on this observation, we conjecture that suppressing recording device characteristics in the input spectrogram is the most effective. We propose a frequency-centering operation for spectrograms that improves the ASC performance on unseen recording devices on average by up to 18.2 percentage points.
Domain shift is considered a challenge in machine learning as it causes significant degradation of model performance. In the Acoustic Scene Classification task (ASC), domain shift is mainly caused by different recording devices. Several studies have already targeted domain generalization to improve the performance of ASC models on unseen domains, such as new devices. Recently, the Controllable Gate Adapter ConGater has been proposed in Natural Language Processing to address the biased training data problem. ConGater allows controlling the debiasing process at inference time. ConGater's main advantage is the continuous and selective debiasing of a trained model, during inference. In this work, we adapt ConGater to the audio spectrogram transformer for an acoustic scene classification task. We show that ConGater can be used to selectively adapt the learned representations to be invariant to device domain shifts such as recording devices. Our analysis shows that ConGater can progressively remove device information from the learned representations and improve the model generalization, especially under domain shift conditions (e.g. unseen devices). We show that information removal can be extended to both device and location domain. Finally, we demonstrate ConGater's ability to enhance specific device performance without further training.
Recent efforts have been made on acoustic scene classification in the audio signal processing community. In contrast, few studies have been conducted on acoustic scene clustering, which is a newly emerging problem. Acoustic scene clustering aims at merging the audio recordings of the same class of acoustic scene into a single cluster without using prior information and training classifiers. In this study, we propose a method for acoustic scene clustering that jointly optimizes the procedures of feature learning and clustering iteration. In the proposed method, the learned feature is a deep embedding that is extracted from a deep convolutional neural network (CNN), while the clustering algorithm is the agglomerative hierarchical clustering (AHC). We formulate a unified loss function for integrating and optimizing these two procedures. Various features and methods are compared. The experimental results demonstrate that the proposed method outperforms other unsupervised methods in terms of the normalized mutual information and the clustering accuracy. In addition, the deep embedding outperforms many state-of-the-art features.
We present a work on low-complexity acoustic scene classification (ASC) with multiple devices, namely the subtask A of Task 1 of the DCASE2021 challenge. This subtask focuses on classifying audio samples of multiple devices with a low-complexity model, where two main difficulties need to be overcome. First, the audio samples are recorded by different devices, and there is mismatch of recording devices in audio samples. We reduce the negative impact of the mismatch of recording devices by using some effective strategies, including data augmentation (e.g., mix-up, spectrum correction, pitch shift), usages of multi-patch network structure and channel attention. Second, the model size should be smaller than a threshold (e.g., 128 KB required by the DCASE2021 challenge). To meet this condition, we adopt a ResNet with both depthwise separable convolution and channel attention as the backbone network, and perform model compression. In summary, we propose a low-complexity ASC method using data augmentation and a lightweight ResNet. Evaluated on the official development and evaluation datasets, our method obtains classification accuracy scores of 71.6% and 66.7%, respectively; and obtains Log-loss scores of 1.038 and 1.136, respectively. Our final model size is 110.3 KB which is smaller than the maximum of 128 KB.
Lifelong audio feature extraction involves learning new sound classes incrementally, which is essential for adapting to new data distributions over time. However, optimizing the model only on new data can lead to catastrophic forgetting of previously learned tasks, which undermines the model's ability to perform well over the long term. This paper introduces a new approach to continual audio representation learning called DeCoR. Unlike other methods that store previous data, features, or models, DeCoR indirectly distills knowledge from an earlier model to the latest by predicting quantization indices from a delayed codebook. We demonstrate that DeCoR improves acoustic scene classification accuracy and integrates well with continual self-supervised representation learning. Our approach introduces minimal storage and computation overhead, making it a lightweight and efficient solution for continual learning.
In this technical report, a low-complexity deep learning system for acoustic scene classification (ASC) is presented. The proposed system comprises two main phases: (Phase I) Training a teacher network; and (Phase II) training a student network using distilled knowledge from the teacher. In the first phase, the teacher, which presents a large footprint model, is trained. After training the teacher, the embeddings, which are the feature map of the second last layer of the teacher, are extracted. In the second phase, the student network, which presents a low complexity model, is trained with the embeddings extracted from the teacher. Our experiments conducted on DCASE 2023 Task 1 Development dataset have fulfilled the requirement of low-complexity and achieved the best classification accuracy of 57.4%, improving DCASE baseline by 14.5%.
The ability to generalize to a wide range of recording devices is a crucial performance factor for audio classification models. The characteristics of different types of microphones introduce distributional shifts in the digitized audio signals due to their varying frequency responses. If this domain shift is not taken into account during training, the model's performance could degrade severely when it is applied to signals recorded by unseen devices. In particular, training a model on audio signals recorded with a small number of different microphones can make generalization to unseen devices difficult. To tackle this problem, we convolve audio signals in the training set with pre-recorded device impulse responses (DIRs) to artificially increase the diversity of recording devices. We systematically study the effect of DIR augmentation on the task of Acoustic Scene Classification using CNNs and Audio Spectrogram Transformers. The results show that DIR augmentation in isolation performs similarly to the state-of-the-art method Freq-MixStyle. However, we also show that DIR augmentation and Freq-MixStyle are complementary, achieving a new state-of-the-art performance on signals recorded by devices unseen during training.
Convolutional neural networks (CNNs) are commonplace in high-performing solutions to many real-world problems, such as audio classification. CNNs have many parameters and filters, with some having a larger impact on the performance than others. This means that networks may contain many unnecessary filters, increasing a CNN's computation and memory requirements while providing limited performance benefits. To make CNNs more efficient, we propose a pruning framework that eliminates filters with the highest "commonality". We measure this commonality using the graph-theoretic concept of "centrality". We hypothesise that a filter with a high centrality should be eliminated as it represents commonality and can be replaced by other filters without affecting the performance of a network much. An experimental evaluation of the proposed framework is performed on acoustic scene classification and audio tagging. On the DCASE 2021 Task 1A baseline network, our proposed method reduces computations per inference by 71\% with 50\% fewer parameters at less than a two percentage point drop in accuracy compared to the original network. For large-scale CNNs such as PANNs designed for audio tagging, our method reduces 24\% computations per inference with 41\% fewer parameters at a slight improvement in performance.
Recent advances in using language models to obtain cross-modal audio-text representations have overcome the limitations of conventional training approaches that use predefined labels. This has allowed the community to make progress in tasks like zero-shot classification, which would otherwise not be possible. However, learning such representations requires a large amount of human-annotated audio-text pairs. In this paper, we study unsupervised approaches to improve the learning framework of such representations with unpaired text and audio. We explore domain-unspecific and domain-specific curation methods to create audio-text pairs that we use to further improve the model. We also show that when domain-specific curation is used in conjunction with a soft-labeled contrastive loss, we are able to obtain significant improvement in terms of zero-shot classification performance on downstream sound event classification or acoustic scene classification tasks.
In this paper, we propose a method for incremental learning of two distinct tasks over time: acoustic scene classification (ASC) and audio tagging (AT). We use a simple convolutional neural network (CNN) model as an incremental learner to solve the tasks. Generally, incremental learning methods catastrophically forget the previous task when sequentially trained on a new task. To alleviate this problem, we propose independent learning and knowledge distillation (KD) between the timesteps in learning. Experiments are performed on TUT 2016/2017 dataset, containing 4 acoustic scene classes and 25 sound event classes. The proposed incremental learner first solves the ASC task with an accuracy of 94.0%. Next, it learns to solve the AT task with an F1 score of 54.4%. At the same time, its performance on the previous ASC task decreases only by 5.1 percentage points due to the additional learning of the AT task.
The real-time processing of time series signals is a critical issue for many real-life applications. The idea of real-time processing is especially important in audio domain as the human perception of sound is sensitive to any kind of disturbance in perceived signals, especially the lag between auditory and visual modalities. The rise of deep learning (DL) models complicated the landscape of signal processing. Although they often have superior quality compared to standard DSP methods, this advantage is diminished by higher latency. In this work we propose novel method for minimization of inference time latency and memory consumption, called Short-Term Memory Convolution (STMC) and its transposed counterpart. The main advantage of STMC is the low latency comparable to long short-term memory (LSTM) networks. Furthermore, the training of STMC-based models is faster and more stable as the method is based solely on convolutional neural networks (CNNs). In this study we demonstrate an application of this solution to a U-Net model for a speech separation task and GhostNet model in acoustic scene classification (ASC) task. In case of speech separation we achieved a 5-fold reduction in inference time and a 2-fold reduction in latency without affecting the output quality. The inference time for ASC task was up to 4 times faster while preserving the original accuracy.
Pattern recognition from audio signals is an active research topic encompassing audio tagging, acoustic scene classification, music classification, and other areas. Spectrogram and mel-frequency cepstral coefficients (MFCC) are among the most commonly used features for audio signal analysis and classification. Recently, deep convolutional neural networks (CNN) have been successfully used for audio classification problems using spectrogram-based 2D features. In this paper, we present SpectNet, an integrated front-end layer that extracts spectrogram features within a CNN architecture that can be used for audio pattern recognition tasks. The front-end layer utilizes learnable gammatone filters that are initialized using mel-scale filters. The proposed layer outputs a 2D spectrogram image which can be fed into a 2D CNN for classification. The parameters of the entire network, including the front-end filterbank, can be updated via back-propagation. This training scheme allows for fine-tuning the spectrogram-image features according to the target audio dataset. The proposed method is evaluated in two different audio signal classification tasks: heart sound anomaly detection and acoustic scene classification. The proposed method shows a significant 1.02\% improvement in MACC for the heart sound classification task and 2.11\% improvement in accuracy for the acoustic scene classification task compared to the classical spectrogram image features. The source code of our experiments can be found at \url{https://github.com/mHealthBuet/SpectNet}
This paper describes a pipeline for collecting acoustic scene data by using crowdsourcing. The detailed process of crowdsourcing is explained, including planning, validation criteria, and actual user interfaces. As a result of data collection, we present CochlScene, a novel dataset for acoustic scene classification. Our dataset consists of 76k samples collected from 831 participants in 13 acoustic scenes. We also propose a manual data split of training, validation, and test sets to increase the reliability of the evaluation results. Finally, we provide a baseline system for future research.
Convolution neural networks (CNNs) have shown great success in various applications. However, the computational complexity and memory storage of CNNs is a bottleneck for their deployment on resource-constrained devices. Recent efforts towards reducing the computation cost and the memory overhead of CNNs involve similarity-based passive filter pruning methods. Similarity-based passive filter pruning methods compute a pairwise similarity matrix for the filters and eliminate a few similar filters to obtain a small pruned CNN. However, the computational complexity of computing the pairwise similarity matrix is high, particularly when a convolutional layer has many filters. To reduce the computational complexity in obtaining the pairwise similarity matrix, we propose to use an efficient method where the complete pairwise similarity matrix is approximated from only a few of its columns by using a Nystr\"om approximation method. The proposed efficient similarity-based passive filter pruning method is 3 times faster and gives same accuracy at the same reduction in computations for CNNs compared to that of the similarity-based pruning method that computes a complete pairwise similarity matrix. Apart from this, the proposed efficient similarity-based pruning method performs similarly or better than the existing norm-based pruning methods. The efficacy of the proposed pruning method is evaluated on CNNs such as DCASE 2021 Task 1A baseline network and a VGGish network designed for acoustic scene classification.
Most existing deep learning-based acoustic scene classification (ASC) approaches directly utilize representations extracted from spectrograms to identify target scenes. However, these approaches pay little attention to the audio events occurring in the scene despite they provide crucial semantic information. This paper conducts the first study that investigates whether real-life acoustic scenes can be reliably recognized based only on the features that describe a limited number of audio events. To model the task-specific relationships between coarse-grained acoustic scenes and fine-grained audio events, we propose an event relational graph representation learning (ERGL) framework for ASC. Specifically, ERGL learns a graph representation of an acoustic scene from the input audio, where the embedding of each event is treated as a node, while the relationship cues derived from each pair of event embeddings are described by a learned multidimensional edge feature. Experiments on a polyphonic acoustic scene dataset show that the proposed ERGL achieves competitive performance on ASC by using only a limited number of embeddings of audio events without any data augmentations. The validity of the proposed ERGL framework proves the feasibility of recognizing diverse acoustic scenes based on the event relational graph. Our code is available on our homepage (https://github.com/Yuanbo2020/ERGL).
In this paper, we present a comprehensive analysis of Acoustic Scene Classification (ASC), the task of identifying the scene of an audio recording from its acoustic signature. In particular, we firstly propose an inception-based and low footprint ASC model, referred to as the ASC baseline. The proposed ASC baseline is then compared with benchmark and high-complexity network architectures of MobileNetV1, MobileNetV2, VGG16, VGG19, ResNet50V2, ResNet152V2, DenseNet121, DenseNet201, and Xception. Next, we improve the ASC baseline by proposing a novel deep neural network architecture which leverages residual-inception architectures and multiple kernels. Given the novel residual-inception (NRI) model, we further evaluate the trade off between the model complexity and the model accuracy performance. Finally, we evaluate whether sound events occurring in a sound scene recording can help to improve ASC accuracy, then indicate how a sound scene context is well presented by combining both sound scene and sound event information. We conduct extensive experiments on various ASC datasets, including Crowded Scenes, IEEE AASP Challenge on Detection and Classification of Acoustic Scenes and Events (DCASE) 2018 Task 1A and 1B, 2019 Task 1A and 1B, 2020 Task 1A, 2021 Task 1A, 2022 Task 1. The experimental results on several different ASC challenges highlight two main achievements; the first is to propose robust, general, and low complexity ASC systems which are suitable for real-life applications on a wide range of edge devices and mobiles; the second is to propose an effective visualization method for comprehensively presenting a sound scene context.
Sound event detection (SED) and Acoustic scene classification (ASC) are two widely researched audio tasks that constitute an important part of research on acoustic scene analysis. Considering shared information between sound events and acoustic scenes, performing both tasks jointly is a natural part of a complex machine listening system. In this paper, we investigate the usefulness of several spatial audio features in training a joint deep neural network (DNN) model performing SED and ASC. Experiments are performed for two different datasets containing binaural recordings and synchronous sound event and acoustic scene labels to analyse the differences between performing SED and ASC separately or jointly. The presented results show that the use of specific binaural features, mainly the Generalized Cross Correlation with Phase Transform (GCC-phat) and sines and cosines of phase differences, result in a better performing model in both separate and joint tasks as compared with baseline methods based on logmel energies only.
This technical report describes the SurreyAudioTeam22s submission for DCASE 2022 ASC Task 1, Low-Complexity Acoustic Scene Classification (ASC). The task has two rules, (a) the ASC framework should have maximum 128K parameters, and (b) there should be a maximum of 30 millions multiply-accumulate operations (MACs) per inference. In this report, we present low-complexity systems for ASC that follow the rules intended for the task.
This paper presents a low-complexity framework for acoustic scene classification (ASC). Most of the frameworks designed for ASC use convolutional neural networks (CNNs) due to their learning ability and improved performance compared to hand-engineered features. However, CNNs are resource hungry due to their large size and high computational complexity. Therefore, CNNs are difficult to deploy on resource constrained devices. This paper addresses the problem of reducing the computational complexity and memory requirement in CNNs. We propose a low-complexity CNN architecture, and apply pruning and quantization to further reduce the parameters and memory. We then propose an ensemble framework that combines various low-complexity CNNs to improve the overall performance. An experimental evaluation of the proposed framework is performed on the publicly available DCASE 2022 Task 1 that focuses on ASC. The proposed ensemble framework has approximately 60K parameters, requires 19M multiply-accumulate operations and improves the performance by approximately 2-4 percentage points compared to the DCASE 2022 Task 1 baseline network.