Following the successful application of vision transformers in multiple computer vision tasks, these models have drawn the attention of the signal processing community. This is because signals are often represented as spectrograms (e.g. through Discrete Fourier Transform) which can be directly provided as input to vision transformers. However, naively applying transformers to spectrograms is suboptimal. Since the axes represent distinct dimensions, i.e. frequency and time, we argue that a better approach is to separate the attention dedicated to each axis. To this end, we propose the Separable Transformer (SepTr), an architecture that employs two transformer blocks in a sequential manner, the first attending to tokens within the same time interval, and the second attending to tokens within the same frequency bin. We conduct experiments on three benchmark data sets, showing that our separable architecture outperforms conventional vision transformers and other state-of-the-art methods. Unlike standard transformers, SepTr linearly scales the number of trainable parameters with the input size, thus having a lower memory footprint. Our code is available as open source at https://github.com/ristea/septr.
It is easier to hear birds than see them, however, they still play an essential role in nature and they are excellent indicators of deteriorating environmental quality and pollution. Recent advances in Machine Learning and Convolutional Neural Networks allow us to detect and classify bird sounds, by doing this, we can assist researchers in monitoring the status and trends of bird populations and biodiversity in ecosystems. We propose a sound detection and classification pipeline for analyzing complex soundscape recordings and identify birdcalls in the background. Our pipeline learns from weak labels, classifies fine-grained bird vocalizations in the wild, and is robust against background sounds (e.g., airplanes, rain, etc). Our solution achieved 10th place of 816 teams at the BirdCLEF 2021 Challenge hosted on Kaggle.
Audio classification is an important task of mapping audio samples into their corresponding labels. Recently, the transformer model with self-attention mechanisms has been adopted in this field. However, existing audio transformers require large GPU memories and long training time, meanwhile relying on pretrained vision models to achieve high performance, which limits the model's scalability in audio tasks. To combat these problems, we introduce HTS-AT: an audio transformer with a hierarchical structure to reduce the model size and training time. It is further combined with a token-semantic module to map final outputs into class featuremaps, thus enabling the model for the audio event detection (i.e. localization in time). We evaluate HTS-AT on three datasets of audio classification where it achieves new state-of-the-art (SOTA) results on AudioSet and ESC-50, and equals the SOTA on Speech Command V2. It also achieves better performance in event localization than the previous CNN-based models. Moreover, HTS-AT requires only 35% model parameters and 15% training time of the previous audio transformer. These results demonstrate the high performance and high efficiency of HTS-AT.
Transformers in their common form are inherently limited to operate on whole token sequences rather than on one token at a time. Consequently, their use during online inference on time-series data entails considerable redundancy due to the overlap in successive token sequences. In this work, we propose novel formulations of the Scaled Dot-Product Attention, which enable Transformers to perform efficient online token-by-token inference on a continual input stream. Importantly, our modifications are purely to the order of computations, while the outputs and learned weights are identical to those of the original Transformer Encoder. We validate our Continual Transformer Encoder with experiments on the THUMOS14, TVSeries and GTZAN datasets with remarkable results: Our Continual one- and two-block architectures reduce the floating point operations per prediction by up to 63x and 2.6x, respectively, while retaining predictive performance.
Classification of long sequential data is an important Machine Learning task and appears in many application scenarios. Recurrent Neural Networks, Transformers, and Convolutional Neural Networks are three major techniques for learning from sequential data. Among these methods, Temporal Convolutional Networks (TCNs) which are scalable to very long sequences have achieved remarkable progress in time series regression. However, the performance of TCNs for sequence classification is not satisfactory because they use a skewed connection protocol and output classes at the last position. Such asymmetry restricts their performance for classification which depends on the whole sequence. In this work, we propose a symmetric multi-scale architecture called Circular Dilated Convolutional Neural Network (CDIL-CNN), where every position has an equal chance to receive information from other positions at the previous layers. Our model gives classification logits in all positions, and we can apply a simple ensemble learning to achieve a better decision. We have tested CDIL-CNN on various long sequential datasets. The experimental results show that our method has superior performance over many state-of-the-art approaches.
Machines that can represent and describe environmental soundscapes have practical potential, e.g., for audio tagging and captioning systems. Prevailing learning paradigms have been relying on parallel audio-text data, which is, however, scarcely available on the web. We propose VIP-ANT that induces \textbf{A}udio-\textbf{T}ext alignment without using any parallel audio-text data. Our key idea is to share the image modality between bi-modal image-text representations and bi-modal image-audio representations; the image modality functions as a pivot and connects audio and text in a tri-modal embedding space implicitly. In a difficult zero-shot setting with no paired audio-text data, our model demonstrates state-of-the-art zero-shot performance on the ESC50 and US8K audio classification tasks, and even surpasses the supervised state of the art for Clotho caption retrieval (with audio queries) by 2.2\% R@1. We further investigate cases of minimal audio-text supervision, finding that, e.g., just a few hundred supervised audio-text pairs increase the zero-shot audio classification accuracy by 8\% on US8K. However, to match human parity on some zero-shot tasks, our empirical scaling experiments suggest that we would need about $2^{21} \approx 2M$ supervised audio-caption pairs. Our work opens up new avenues for learning audio-text connections with little to no parallel audio-text data.
Voice Activity Detection (VAD) is a fundamental preprocessing step in automatic speech recognition. This is especially true within the broadcast industry where a wide variety of audio materials and recording conditions are encountered. Based on previous studies which indicate that xvector embeddings can be applied to a diverse set of audio classification tasks, we investigate the suitability of x-vectors in discriminating speech from noise. We find that the proposed x-vector based VAD system achieves the best reported score in detecting clean speech on AVA-Speech, whilst retaining robust VAD performance in the presence of noise and music. Furthermore, we integrate the x-vector based VAD system into an existing STT pipeline and compare its performance on multiple broadcast datasets against a baseline system with WebRTC VAD. Crucially, our proposed x-vector based VAD improves the accuracy of STT transcription on real-world broadcast audio
Audio representations for music information retrieval are typically learned via supervised learning in a task-specific fashion. Although effective at producing state-of-the-art results, this scheme lacks flexibility with respect to the range of applications a model can have and requires extensively annotated datasets. In this work, we pose the question of whether it may be possible to exploit weakly aligned text as the only supervisory signal to learn general-purpose music audio representations. To address this question, we design a multimodal architecture for music and language pre-training (MuLaP) optimised via a set of proxy tasks. Weak supervision is provided in the form of noisy natural language descriptions conveying the overall musical content of the track. After pre-training, we transfer the audio backbone of the model to a set of music audio classification and regression tasks. We demonstrate the usefulness of our approach by comparing the performance of audio representations produced by the same audio backbone with different training strategies and show that our pre-training method consistently achieves comparable or higher scores on all tasks and datasets considered. Our experiments also confirm that MuLaP effectively leverages audio-caption pairs to learn representations that are competitive with audio-only and cross-modal self-supervised methods in the literature.
The recent success of the generative model shows that leveraging the multi-modal embedding space can manipulate an image using text information. However, manipulating an image with other sources rather than text, such as sound, is not easy due to the dynamic characteristics of the sources. Especially, sound can convey vivid emotions and dynamic expressions of the real world. Here, we propose a framework that directly encodes sound into the multi-modal (image-text) embedding space and manipulates an image from the space. Our audio encoder is trained to produce a latent representation from an audio input, which is forced to be aligned with image and text representations in the multi-modal embedding space. We use a direct latent optimization method based on aligned embeddings for sound-guided image manipulation. We also show that our method can mix text and audio modalities, which enrich the variety of the image modification. We verify the effectiveness of our sound-guided image manipulation quantitatively and qualitatively. We also show that our method can mix different modalities, i.e., text and audio, which enrich the variety of the image modification. The experiments on zero-shot audio classification and semantic-level image classification show that our proposed model outperforms other text and sound-guided state-of-the-art methods.
We present CrissCross, a self-supervised framework for learning audio-visual representations. A novel notion is introduced in our framework whereby in addition to learning the intra-modal and standard 'synchronous' cross-modal relations, CrissCross also learns 'asynchronous' cross-modal relationships. We perform in-depth studies showing that by relaxing the temporal synchronicity between the audio and visual modalities, the network learns strong generalized representations useful for a variety of downstream tasks. To pretrain our proposed solution, we use 3 different datasets with varying sizes, Kinetics-Sound, Kinetics400, and AudioSet. The learned representations are evaluated on a number of downstream tasks namely action recognition, sound classification, and action retrieval. Our experiments show that CrissCross either outperforms or achieves performances on par with the current state-of-the-art self-supervised methods on action recognition and action retrieval with UCF101 and HMDB51, as well as sound classification with ESC50 and DCASE. Moreover, CrissCross outperforms fully-supervised pretraining while pretrained on Kinetics-Sound. The codes and pretrained models are available on the project website.
Recently, neural networks based purely on self-attention, such as the Vision Transformer (ViT), have been shown to outperform deep learning models constructed with convolutional neural networks (CNNs) on various vision tasks, thus extending the success of Transformers, which were originally developed for language processing, to the vision domain. A recent study showed that a similar methodology can also be applied to the audio domain. Specifically, the Audio Spectrogram Transformer (AST) achieves state-of-the-art results on various audio classification benchmarks. However, pure Transformer models tend to require more training data compared to CNNs, and the success of the AST relies on supervised pretraining that requires a large amount of labeled data and a complex training pipeline, thus limiting the practical usage of AST. This paper focuses on audio and speech classification, and aims to reduce the need for large amounts of labeled data for AST by leveraging self-supervised learning using unlabeled data. Specifically, we propose to pretrain the AST model with joint discriminative and generative masked spectrogram patch modeling (MSPM) using unlabeled audio from AudioSet and Librispeech. We evaluate our pretrained models on both audio and speech classification tasks including audio event classification, keyword spotting, emotion recognition, and speaker identification. The proposed self-supervised framework significantly boosts AST performance on all tasks, with an average improvement of 60.9%, leading to similar or even better results than a supervised pretrained AST. To the best of our knowledge, it is the first patch-based self-supervised learning framework in the audio and speech domain, and also the first self-supervised learning framework for AST.
Transformers have revolutionized the world of deep learning, specially in the field of natural language processing. Recently, the Audio Spectrogram Transformer (AST) was proposed for audio classification, leading to state of the art results in several datasets. However, in order for ASTs to outperform CNNs, pretraining with ImageNet is needed. In this paper, we study one component of the AST, the positional encoding, and propose several variants to improve the performance of ASTs trained from scratch, without ImageNet pretraining. Our best model, which incorporates conditional positional encodings, significantly improves performance on Audioset and ESC-50 compared to the original AST.
Hierarchical taxonomies are common in many contexts, and they are a very natural structure humans use to organise information. In machine learning, the family of methods that use the 'extra' information is called hierarchical classification. However, applied to audio classification, this remains relatively unexplored. Here we focus on how to integrate the hierarchical information of a problem to learn embeddings representative of the hierarchical relationships. Previously, triplet loss has been proposed to address this problem, however it presents some issues like requiring the careful construction of the triplets, and being limited in the extent of hierarchical information it uses at each iteration. In this work we propose a rank based loss function that uses hierarchical information and translates this into a rank ordering of target distances between the examples. We show that rank based loss is suitable to learn hierarchical representations of the data. By testing on unseen fine level classes we show that this method is also capable of learning hierarchically correct representations of the new classes. Rank based loss has two promising aspects, it is generalisable to hierarchies with any number of levels, and is capable of dealing with data with incomplete hierarchical labels.
The great success of transformer-based models in natural language processing (NLP) has led to various attempts at adapting these architectures to other domains such as vision and audio. Recent work has shown that transformers can outperform Convolutional Neural Networks (CNNs) on vision and audio tasks. However, one of the main shortcomings of transformer models, compared to the well-established CNNs, is the computational complexity. In transformers, the compute and memory complexity is known to grow quadratically with the input length. Therefore, there has been extensive work on optimizing transformers, but often at the cost of degrading predictive performance. In this work, we propose a novel method to optimize and regularize transformers on audio spectrograms. Our proposed models achieve a new state-of-the-art performance on Audioset and can be trained on a single consumer-grade GPU. Furthermore, we propose a transformer model that outperforms CNNs in terms of both performance and training speed. Source code: https://github.com/kkoutini/PaSST
Deep learning has celebrated resounding successes in many application areas of relevance to the Internet of Things (IoT), such as computer vision and machine listening. These technologies must ultimately be brought directly to the edge to fully harness the power of deep learning for the IoT. The obvious challenge is that deep learning techniques can only be implemented on strictly resource-constrained edge devices if the models are radically downsized. This task relies on different model compression techniques, such as network pruning, quantization, and the recent advancement of XNOR-Net. This study examines the suitability of these techniques for audio classification on microcontrollers. We present an application of XNOR-Net for end-to-end raw audio classification and a comprehensive empirical study comparing this approach with pruning-and-quantization methods. We show that raw audio classification with XNOR yields comparable performance to regular full precision networks for small numbers of classes while reducing memory requirements 32-fold and computation requirements 58-fold. However, as the number of classes increases significantly, performance degrades, and pruning-and-quantization based compression techniques take over as the preferred technique being able to satisfy the same space constraints but requiring approximately 8x more computation. We show that these insights are consistent between raw audio classification and image classification using standard benchmark sets. To the best of our knowledge, this is the first study to apply XNOR to end-to-end audio classification and evaluate it in the context of alternative techniques. All codes are publicly available on GitHub.
Exponential families are widely used in machine learning, including many distributions in continuous and discrete domains (e.g., Gaussian, Dirichlet, Poisson, and categorical distributions via the softmax transformation). Distributions in each of these families have fixed support. In contrast, for finite domains, recent work on sparse alternatives to softmax (e.g., sparsemax, $\alpha$-entmax, and fusedmax), has led to distributions with varying support. This paper develops sparse alternatives to continuous distributions, based on several technical contributions: First, we define $\Omega$-regularized prediction maps and Fenchel-Young losses for arbitrary domains (possibly countably infinite or continuous). For linearly parametrized families, we show that minimization of Fenchel-Young losses is equivalent to moment matching of the statistics, generalizing a fundamental property of exponential families. When $\Omega$ is a Tsallis negentropy with parameter $\alpha$, we obtain ``deformed exponential families,'' which include $\alpha$-entmax and sparsemax ($\alpha=2$) as particular cases. For quadratic energy functions, the resulting densities are $\beta$-Gaussians, an instance of elliptical distributions that contain as particular cases the Gaussian, biweight, triweight, and Epanechnikov densities, and for which we derive closed-form expressions for the variance, Tsallis entropy, and Fenchel-Young loss. When $\Omega$ is a total variation or Sobolev regularizer, we obtain a continuous version of the fusedmax. Finally, we introduce continuous-domain attention mechanisms, deriving efficient gradient backpropagation algorithms for $\alpha \in \{1, 4/3, 3/2, 2\}$. Using these algorithms, we demonstrate our sparse continuous distributions for attention-based audio classification and visual question answering, showing that they allow attending to time intervals and compact regions.
Local explanation methods such as LIME have become popular in MIR as tools for generating post-hoc, model-agnostic explanations of a model's classification decisions. The basic idea is to identify a small set of human-understandable features of the classified example that are most influential on the classifier's prediction. These are then presented as an explanation. Evaluation of such explanations in publications often resorts to accepting what matches the expectation of a human without actually being able to verify if what the explanation shows is what really caused the model's prediction. This paper reports on targeted investigations where we try to get more insight into the actual veracity of LIME's explanations in an audio classification task. We deliberately design adversarial examples for the classifier, in a way that gives us knowledge about which parts of the input are potentially responsible for the model's (wrong) prediction. Asking LIME to explain the predictions for these adversaries permits us to study whether local explanations do indeed detect these regions of interest. We also look at whether LIME is more successful in finding perturbations that are more prominent and easily noticeable for a human. Our results suggest that LIME does not necessarily manage to identify the most relevant input features and hence it remains unclear whether explanations are useful or even misleading.
Federated Learning is a distributed machine learning paradigm dealing with decentralized and personal datasets. Since data reside on devices like smartphones and virtual assistants, labeling is entrusted to the clients, or labels are extracted in an automated way. Specifically, in the case of audio data, acquiring semantic annotations can be prohibitively expensive and time-consuming. As a result, an abundance of audio data remains unlabeled and unexploited on users' devices. Most existing federated learning approaches focus on supervised learning without harnessing the unlabeled data. In this work, we study the problem of semi-supervised learning of audio models via self-training in conjunction with federated learning. We propose FedSTAR to exploit large-scale on-device unlabeled data to improve the generalization of audio recognition models. We further demonstrate that self-supervised pre-trained models can accelerate the training of on-device models, significantly improving convergence to within fewer training rounds. We conduct experiments on diverse public audio classification datasets and investigate the performance of our models under varying percentages of labeled and unlabeled data. Notably, we show that with as little as 3% labeled data available, FedSTAR on average can improve the recognition rate by 13.28% compared to the fully supervised federated model.
Humans perceive the world by concurrently processing and fusing high-dimensional inputs from multiple modalities such as vision and audio. Machine perception models, in stark contrast, are typically modality-specific and optimised for unimodal benchmarks, and hence late-stage fusion of final representations or predictions from each modality (`late-fusion') is still a dominant paradigm for multimodal video classification. Instead, we introduce a novel transformer based architecture that uses `fusion bottlenecks' for modality fusion at multiple layers. Compared to traditional pairwise self-attention, our model forces information between different modalities to pass through a small number of bottleneck latents, requiring the model to collate and condense the most relevant information in each modality and only share what is necessary. We find that such a strategy improves fusion performance, at the same time reducing computational cost. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple audio-visual classification benchmarks including Audioset, Epic-Kitchens and VGGSound. All code and models will be released.
Sound event localization and detection consists of two subtasks which are sound event detection and direction-of-arrival estimation. While sound event detection mainly relies on time-frequency patterns to distinguish different sound classes, direction-of-arrival estimation uses magnitude or phase differences between microphones to estimate source directions. Therefore, it is often difficult to jointly train these two subtasks simultaneously. We propose a novel feature called spatial cue-augmented log-spectrogram (SALSA) with exact time-frequency mapping between the signal power and the source direction-of-arrival. The feature includes multichannel log-spectrograms stacked along with the estimated direct-to-reverberant ratio and a normalized version of the principal eigenvector of the spatial covariance matrix at each time-frequency bin on the spectrograms. Experimental results on the DCASE 2021 dataset for sound event localization and detection with directional interference showed that the deep learning-based models trained on this new feature outperformed the DCASE challenge baseline by a large margin. We combined several models with slightly different architectures that were trained on the new feature to further improve the system performances for the DCASE sound event localization and detection challenge.