Sound event localization and detection (SELD) systems estimate both the direction-of-arrival (DOA) and class of sound sources over time. In the DCASE 2022 SELD Challenge (Task 3), models are designed to operate in a 4-channel setting. While beneficial to further the development of SELD systems using a multichannel recording setup such as first-order Ambisonics (FOA), most consumer electronics devices rarely are able to record using more than two channels. For this reason, in this work we investigate the performance of the DCASE 2022 SELD baseline model using three audio input representations: FOA, binaural, and stereo. We perform a novel comparative analysis illustrating the effect of these audio input representations on SELD performance. Crucially, we show that binaural and stereo (i.e. 2-channel) audio-based SELD models are still able to localize and detect sound sources laterally quite well, despite overall performance degrading as less audio information is provided. Further, we segment our analysis by scenes containing varying degrees of sound source polyphony to better understand the effect of audio input representation on localization and detection performance as scene conditions become increasingly complex.
Localizing a moving sound source in the real world involves determining its direction-of-arrival (DOA) and distance relative to a microphone. Advancements in DOA estimation have been facilitated by data-driven methods optimized with large open-source datasets with microphone array recordings in diverse environments. In contrast, estimating a sound source's distance remains understudied. Existing approaches assume recordings by non-coincident microphones to use methods that are susceptible to differences in room reverberation. We present a CRNN able to estimate the distance of moving sound sources across multiple datasets featuring diverse rooms, outperforming a recently-published approach. We also characterize our model's performance as a function of sound source distance and different training losses. This analysis reveals optimal training using a loss that weighs model errors as an inverse function of the sound source true distance. Our study is the first to demonstrate that sound source distance estimation can be performed across diverse acoustic conditions using deep learning.
Finding the right sound effects (SFX) to match moments in a video is a difficult and time-consuming task, and relies heavily on the quality and completeness of text metadata. Retrieving high-quality (HQ) SFX using a video frame directly as the query is an attractive alternative, removing the reliance on text metadata and providing a low barrier to entry for non-experts. Due to the lack of HQ audio-visual training data, previous work on audio-visual retrieval relies on YouTube (in-the-wild) videos of varied quality for training, where the audio is often noisy and the video of amateur quality. As such it is unclear whether these systems would generalize to the task of matching HQ audio to production-quality video. To address this, we propose a multimodal framework for recommending HQ SFX given a video frame by (1) leveraging large language models and foundational vision-language models to bridge HQ audio and video to create audio-visual pairs, resulting in a highly scalable automatic audio-visual data curation pipeline; and (2) using pre-trained audio and visual encoders to train a contrastive learning-based retrieval system. We show that our system, trained using our automatic data curation pipeline, significantly outperforms baselines trained on in-the-wild data on the task of HQ SFX retrieval for video. Furthermore, while the baselines fail to generalize to this task, our system generalizes well from clean to in-the-wild data, outperforming the baselines on a dataset of YouTube videos despite only being trained on the HQ audio-visual pairs. A user study confirms that people prefer SFX retrieved by our system over the baseline 67% of the time both for HQ and in-the-wild data. Finally, we present ablations to determine the impact of model and data pipeline design choices on downstream retrieval performance. Please visit our project website to listen to and view our SFX retrieval results.
In the context of environmental sound classification, the adaptability of systems is key: which sound classes are interesting depends on the context and the user's needs. Recent advances in text-to-audio retrieval allow for zero-shot audio classification, but performance compared to supervised models remains limited. This work proposes a multimodal prototypical approach that exploits local audio-text embeddings to provide more relevant answers to audio queries, augmenting the adaptability of sound detection in the wild. We do this by first using text to query a nearby community of audio embeddings that best characterize each query sound, and select the group's centroids as our prototypes. Second, we compare unseen audio to these prototypes for classification. We perform multiple ablation studies to understand the impact of the embedding models and prompts. Our unsupervised approach improves upon the zero-shot state-of-the-art in three sound recognition benchmarks by an average of 12%.
Beat and downbeat tracking models have improved significantly in recent years with the introduction of deep learning methods. However, despite these improvements, several challenges remain. Particularly, the adaptation of available models to underrepresented music traditions in MIR is usually synonymous with collecting and annotating large amounts of data, which is impractical and time-consuming. Transfer learning, data augmentation, and fine-tuning techniques have been used quite successfully in related tasks and are known to alleviate this bottleneck. Furthermore, when studying these music traditions, models are not required to generalize to multiple mainstream music genres but to perform well in more constrained, homogeneous conditions. In this work, we investigate simple yet effective strategies to adapt beat and downbeat tracking models to two different Latin American music traditions and analyze the feasibility of these adaptations in real-world applications concerning the data and computational requirements. Contrary to common belief, our findings show it is possible to achieve good performance by spending just a few minutes annotating a portion of the data and training a model in a standard CPU machine, with the precise amount of resources needed depending on the task and the complexity of the dataset.
Self-supervision methods learn representations by solving pretext tasks that do not require human-generated labels, alleviating the need for time-consuming annotations. These methods have been applied in computer vision, natural language processing, environmental sound analysis, and recently in music information retrieval, e.g. for pitch estimation. Particularly in the context of music, there are few insights about the fragility of these models regarding different distributions of data, and how they could be mitigated. In this paper, we explore these questions by dissecting a self-supervised model for pitch estimation adapted for tempo estimation via rigorous experimentation with synthetic data. Specifically, we study the relationship between the input representation and data distribution for self-supervised tempo estimation.
Most recent work in visual sound source localization relies on semantic audio-visual representations learned in a self-supervised manner, and by design excludes temporal information present in videos. While it proves to be effective for widely used benchmark datasets, the method falls short for challenging scenarios like urban traffic. This work introduces temporal context into the state-of-the-art methods for sound source localization in urban scenes using optical flow as a means to encode motion information. An analysis of the strengths and weaknesses of our methods helps us better understand the problem of visual sound source localization and sheds light on open challenges for audio-visual scene understanding.
Localizing visual sounds consists on locating the position of objects that emit sound within an image. It is a growing research area with potential applications in monitoring natural and urban environments, such as wildlife migration and urban traffic. Previous works are usually evaluated with datasets having mostly a single dominant visible object, and proposed models usually require the introduction of localization modules during training or dedicated sampling strategies, but it remains unclear how these design choices play a role in the adaptability of these methods in more challenging scenarios. In this work, we analyze various model choices for visual sound localization and discuss how their different components affect the model's performance, namely the encoders' architecture, the loss function and the localization strategy. Furthermore, we study the interaction between these decisions, the model performance, and the data, by digging into different evaluation datasets spanning different difficulties and characteristics, and discuss the implications of such decisions in the context of real-world applications. Our code and model weights are open-sourced and made available for further applications.
Audio applications involving environmental sound analysis increasingly use general-purpose audio representations, also known as embeddings, for transfer learning. Recently, Holistic Evaluation of Audio Representations (HEAR) evaluated twenty-nine embedding models on nineteen diverse tasks. However, the evaluation's effectiveness depends on the variation already captured within a given dataset. Therefore, for a given data domain, it is unclear how the representations would be affected by the variations caused by myriad microphones' range and acoustic conditions -- commonly known as channel effects. We aim to extend HEAR to evaluate invariance to channel effects in this work. To accomplish this, we imitate channel effects by injecting perturbations to the audio signal and measure the shift in the new (perturbed) embeddings with three distance measures, making the evaluation domain-dependent but not task-dependent. Combined with the downstream performance, it helps us make a more informed prediction of how robust the embeddings are to the channel effects. We evaluate two embeddings -- YAMNet, and OpenL$^3$ on monophonic (UrbanSound8K) and polyphonic (SONYC UST) datasets. We show that one distance measure does not suffice in such task-independent evaluation. Although Fr\'echet Audio Distance (FAD) correlates with the trend of the performance drop in the downstream task most accurately, we show that we need to study this in conjunction with the other distances to get a clear understanding of the overall effect of the perturbation. In terms of the embedding performance, we find OpenL$^3$ to be more robust to YAMNet, which aligns with the HEAR evaluation.
Soundata is a Python library for loading and working with audio datasets in a standardized way, removing the need for writing custom loaders in every project, and improving reproducibility by providing tools to validate data against a canonical version. It speeds up research pipelines by allowing users to quickly download a dataset, load it into memory in a standardized and reproducible way, validate that the dataset is complete and correct, and more. Soundata is based and inspired on mirdata and design to complement mirdata by working with environmental sound, bioacoustic and speech datasets, among others. Soundata was created to be easy to use, easy to contribute to, and to increase reproducibility and standardize usage of sound datasets in a flexible way.