A fundamental characteristic of audio is its compositional nature. Audio-language models (ALMs) trained using a contrastive approach (e.g., CLAP) that learns a shared representation between audio and language modalities have improved performance in many downstream applications, including zero-shot audio classification, audio retrieval, etc. However, the ability of these models to effectively perform compositional reasoning remains largely unexplored and necessitates additional research. In this paper, we propose CompA, a collection of two expert-annotated benchmarks with a majority of real-world audio samples, to evaluate compositional reasoning in ALMs. Our proposed CompA-order evaluates how well an ALM understands the order or occurrence of acoustic events in audio, and CompA-attribute evaluates attribute binding of acoustic events. An instance from either benchmark consists of two audio-caption pairs, where both audios have the same acoustic events but with different compositions. An ALM is evaluated on how well it matches the right audio to the right caption. Using this benchmark, we first show that current ALMs perform only marginally better than random chance, thereby struggling with compositional reasoning. Next, we propose CompA-CLAP, where we fine-tune CLAP using a novel learning method to improve its compositional reasoning abilities. To train CompA-CLAP, we first propose improvements to contrastive training with composition-aware hard negatives, allowing for more focused training. Next, we propose a novel modular contrastive loss that helps the model learn fine-grained compositional understanding and overcomes the acute scarcity of openly available compositional audios. CompA-CLAP significantly improves over all our baseline models on the CompA benchmark, indicating its superior compositional reasoning capabilities.
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.
Spoken language recognition (SLR) is the task of automatically identifying the language present in a speech signal. Existing SLR models are either too computationally expensive or too large to run effectively on devices with limited resources. For real-world deployment, a model should also gracefully handle unseen languages outside of the target language set, yet prior work has focused on closed-set classification where all input languages are known a-priori. In this paper we address these two limitations: we explore efficient model architectures for SLR based on convolutional networks, and propose a multilabel training strategy to handle non-target languages at inference time. Using the VoxLingua107 dataset, we show that our models obtain competitive results while being orders of magnitude smaller and faster than current state-of-the-art methods, and that our multilabel strategy is more robust to unseen non-target languages compared to multiclass classification.
We propose a self-supervised approach for learning to perform audio source separation in videos based on natural language queries, using only unlabeled video and audio pairs as training data. A key challenge in this task is learning to associate the linguistic description of a sound-emitting object to its visual features and the corresponding components of the audio waveform, all without access to annotations during training. To overcome this challenge, we adapt off-the-shelf vision-language foundation models to provide pseudo-target supervision via two novel loss functions and encourage a stronger alignment between the audio, visual and natural language modalities. During inference, our approach can separate sounds given text, video and audio input, or given text and audio input alone. We demonstrate the effectiveness of our self-supervised approach on three audio-visual separation datasets, including MUSIC, SOLOS and AudioSet, where we outperform state-of-the-art strongly supervised approaches despite not using object detectors or text labels during training.
Multi-modal contrastive learning techniques in the audio-text domain have quickly become a highly active area of research. Most works are evaluated with standard audio retrieval and classification benchmarks assuming that (i) these models are capable of leveraging the rich information contained in natural language, and (ii) current benchmarks are able to capture the nuances of such information. In this work, we show that state-of-the-art audio-text models do not yet really understand natural language, especially contextual concepts such as sequential or concurrent ordering of sound events. Our results suggest that existing benchmarks are not sufficient to assess these models' capabilities to match complex contexts from the audio and text modalities. We propose a Transformer-based architecture and show that, unlike prior work, it is capable of modeling the sequential relationship between sound events in the text and audio, given appropriate benchmark data. We advocate for the collection or generation of additional, diverse, data to allow future research to fully leverage natural language for audio-text modeling.
Consumer-grade music recordings such as those captured by mobile devices typically contain distortions in the form of background noise, reverb, and microphone-induced EQ. This paper presents a deep learning approach to enhance low-quality music recordings by combining (i) an image-to-image translation model for manipulating audio in its mel-spectrogram representation and (ii) a music vocoding model for mapping synthetically generated mel-spectrograms to perceptually realistic waveforms. We find that this approach to music enhancement outperforms baselines which use classical methods for mel-spectrogram inversion and an end-to-end approach directly mapping noisy waveforms to clean waveforms. Additionally, in evaluating the proposed method with a listening test, we analyze the reliability of common audio enhancement evaluation metrics when used in the music domain.
The mood of a song is a highly relevant feature for exploration and recommendation in large collections of music. These collections tend to require automatic methods for predicting such moods. In this work, we show that listening-based features outperform content-based ones when classifying moods: embeddings obtained through matrix factorization of listening data appear to be more informative of a track mood than embeddings based on its audio content. To demonstrate this, we compile a subset of the Million Song Dataset, totalling 67k tracks, with expert annotations of 188 different moods collected from AllMusic. Our results on this novel dataset not only expose the limitations of current audio-based models, but also aim to foster further reproducible research on this timely topic.
The lack of data tends to limit the outcomes of deep learning research, particularly when dealing with end-to-end learning stacks processing raw data such as waveforms. In this study, 1.2M tracks annotated with musical labels are available to train our end-to-end models. This large amount of data allows us to unrestrictedly explore two different design paradigms for music auto-tagging: assumption-free models - using waveforms as input with very small convolutional filters; and models that rely on domain knowledge - log-mel spectrograms with a convolutional neural network designed to learn timbral and temporal features. Our work focuses on studying how these two types of deep architectures perform when datasets of variable size are available for training: the MagnaTagATune (25k songs), the Million Song Dataset (240k songs), and a private dataset of 1.2M songs. Our experiments suggest that music domain assumptions are relevant when not enough training data are available, thus showing how waveform-based models outperform spectrogram-based ones in large-scale data scenarios.
Online audio advertising is a particular form of advertising used abundantly in online music streaming services. In these platforms, which tend to host tens of thousands of unique audio advertisements (ads), providing high quality ads ensures a better user experience and results in longer user engagement. Therefore, the automatic assessment of these ads is an important step toward audio ads ranking and better audio ads creation. In this paper we propose one way to measure the quality of the audio ads using a proxy metric called Long Click Rate (LCR), which is defined by the amount of time a user engages with the follow-up display ad (that is shown while the audio ad is playing) divided by the impressions. We later focus on predicting the audio ad quality using only acoustic features such as harmony, rhythm, and timbre of the audio, extracted from the raw waveform. We discuss how the characteristics of the sound can be connected to concepts such as the clarity of the audio ad message, its trustworthiness, etc. Finally, we propose a new deep learning model for audio ad quality prediction, which outperforms the other discussed models trained on hand-crafted features. To the best of our knowledge, this is the first large-scale audio ad quality prediction study.
An increasing amount of digital music is being published daily. Music streaming services often ingest all available music, but this poses a challenge: how to recommend new artists for which prior knowledge is scarce? In this work we aim to address this so-called cold-start problem by combining text and audio information with user feedback data using deep network architectures. Our method is divided into three steps. First, artist embeddings are learned from biographies by combining semantics, text features, and aggregated usage data. Second, track embeddings are learned from the audio signal and available feedback data. Finally, artist and track embeddings are combined in a multimodal network. Results suggest that both splitting the recommendation problem between feature levels (i.e., artist metadata and audio track), and merging feature embeddings in a multimodal approach improve the accuracy of the recommendations.