Paul Primus

Paul Primus, Khaled Koutini, Gerhard Widmer

This work presents a text-to-audio-retrieval system based on pre-trained text and spectrogram transformers. Our method projects recordings and textual descriptions into a shared audio-caption space in which related examples from different modalities are close. Through a systematic analysis, we examine how each component of the system influences retrieval performance. As a result, we identify two key components that play a crucial role in driving performance: the self-attention-based audio encoder for audio embedding and the utilization of additional human-generated and synthetic data sets during pre-training. We further experimented with augmenting ClothoV2 captions with available keywords to increase their variety; however, this only led to marginal improvements. Our system ranked first in the 2023's DCASE Challenge, and it outperforms the current state of the art on the ClothoV2 benchmark by 5.6 pp. mAP@10.

Paul Primus, Gerhard Widmer

Standard machine learning models for tagging and classifying acoustic signals cannot handle classes that were not seen during training. Zero-Shot (ZS) learning overcomes this restriction by predicting classes based on adaptable class descriptions. This study sets out to investigate the effectiveness of self-attention-based audio embedding architectures for ZS learning. To this end, we compare the very recent patchout spectrogram transformer with two classic convolutional architectures. We evaluate these three architectures on three tasks and on three different benchmark datasets: general-purpose tagging on AudioSet, environmental sound classification on ESC-50, and instrument tagging on OpenMIC. Our results show that the self-attention-based embedding methods outperform both compared convolutional architectures in all of these settings. By designing training and test data accordingly, we observe that prediction performance suffers significantly when the `semantic distance' between training and new test classes is large, an effect that will deserve more detailed investigations.

Paul Primus, Gerhard Widmer

The absence of large labeled datasets remains a significant challenge in many application areas of deep learning. Researchers and practitioners typically resort to transfer learning and data augmentation to alleviate this issue. We study these strategies in the context of audio retrieval with natural language queries (Task 6b of the DCASE 2022 Challenge). Our proposed system uses pre-trained embedding models to project recordings and textual descriptions into a shared audio-caption space in which related examples from different modalities are close. We employ various data augmentation techniques on audio and text inputs and systematically tune their corresponding hyperparameters with sequential model-based optimization. Our results show that the used augmentations strategies reduce overfitting and improve retrieval performance.

Florian Schmid, Paul Primus, Toni Heittola et al.

This paper presents the Low-Complexity Acoustic Scene Classification with Device Information Task of the DCASE 2025 Challenge, along with its baseline system. Continuing the focus on low-complexity models, data efficiency, and device mismatch from previous editions (2022-2024), this year's task introduces a key change: recording device information is now provided at inference time. This enables the development of device-specific models that leverage device characteristics-reflecting real-world deployment scenarios in which a model is designed with awareness of the underlying hardware. The training set matches the 25% subset used in the corresponding DCASE 2024 challenge, with no restrictions on external data use, highlighting transfer learning as a central topic. The baseline achieves 50.72% accuracy with a device-agnostic model, improving to 51.89% when incorporating device-specific fine-tuning. The task attracted 31 submissions from 12 teams, with 11 teams outperforming the baseline. The top-performing submission achieved an accuracy gain of more than 8 percentage points over the baseline on the evaluation set.

Paul Primus, Florian Schmid, Gerhard Widmer

Dual-encoder-based audio retrieval systems are commonly optimized with contrastive learning on a set of matching and mismatching audio-caption pairs. This leads to a shared embedding space in which corresponding items from the two modalities end up close together. Since audio-caption datasets typically only contain matching pairs of recordings and descriptions, it has become common practice to create mismatching pairs by pairing the audio with a caption randomly drawn from the dataset. This is not ideal because the randomly sampled caption could, just by chance, partly or entirely describe the audio recording. However, correspondence information for all possible pairs is costly to annotate and thus typically unavailable; we, therefore, suggest substituting it with estimated correspondences. To this end, we propose a two-staged training procedure in which multiple retrieval models are first trained as usual, i.e., without estimated correspondences. In the second stage, the audio-caption correspondences predicted by these models then serve as prediction targets. We evaluate our method on the ClothoV2 and the AudioCaps benchmark and show that it improves retrieval performance, even in a restricting self-distillation setting where a single model generates and then learns from the estimated correspondences. We further show that our method outperforms the current state of the art by 1.6 pp. mAP@10 on the ClothoV2 benchmark.

Tobias Morocutti, Jonathan Greif, Paul Primus et al.

Spatial semantic segmentation of sound scenes (S5) involves the accurate identification of active sound classes and the precise separation of their sources from complex acoustic mixtures. Conventional systems rely on a two-stage pipeline - audio tagging followed by label-conditioned source separation - but are often constrained by the absence of fine-grained temporal information critical for effective separation. In this work, we address this limitation by introducing a novel approach for S5 that enhances the synergy between the event detection and source separation stages. Our key contributions are threefold. First, we fine-tune a pre-trained Transformer to detect active sound classes. Second, we utilize a separate instance of this fine-tuned Transformer to perform sound event detection (SED), providing the separation module with detailed, time-varying guidance. Third, we implement an iterative refinement mechanism that progressively enhances separation quality by recursively reusing the separator's output from previous iterations. These advancements lead to significant improvements in both audio tagging and source separation performance, as demonstrated by our system's second-place finish in Task 4 of the DCASE Challenge 2025. Our implementation and model checkpoints are available in our GitHub repository: https://github.com/theMoro/dcase25task4 .

Paul Primus, Florian Schmid, Gerhard Widmer

Learning to associate audio with textual descriptions is valuable for a range of tasks, including pretraining, zero-shot classification, audio retrieval, audio captioning, and text-conditioned audio generation. Existing contrastive language-audio pretrained models are typically trained using global, clip-level descriptions, which provide only weak temporal supervision. We hypothesize that CLAP-like language-audio models - particularly, if they are expected to produce frame-level embeddings - can benefit from a stronger temporal supervision. To confirm our hypothesis, we curate a novel dataset of approximately 12,000 audio recordings from Freesound, each annotated with single-sentence free-text descriptions linked to a specific temporal segment in an audio recording. We use large language models to clean these annotations by removing references to non-audible events, transcribed speech, typos, and annotator language bias. We further propose a frame-wise contrastive training strategy that learns to align text descriptions with temporal regions in an audio recording and demonstrate that our model has better temporal text-audio alignment abilities compared to models trained only on global captions when evaluated on the AudioSet Strong benchmark. The dataset and our source code are available on Zenodo and GitHub, respectively.

Paul Primus, Gerhard Widmer

Matching raw audio signals with textual descriptions requires understanding the audio's content and the description's semantics and then drawing connections between the two modalities. This paper investigates a hybrid retrieval system that utilizes audio metadata as an additional clue to understand the content of audio signals before matching them with textual queries. We experimented with metadata often attached to audio recordings, such as keywords and natural-language descriptions, and we investigated late and mid-level fusion strategies to merge audio and metadata. Our hybrid approach with keyword metadata and late fusion improved the retrieval performance over a content-based baseline by 2.36 and 3.69 pp. mAP@10 on the ClothoV2 and AudioCaps benchmarks, respectively.

Paul Primus, Verena Haunschmid, Patrick Praher et al.

Unsupervised anomalous sound detection is concerned with identifying sounds that deviate from what is defined as 'normal', without explicitly specifying the types of anomalies. A significant obstacle is the diversity and rareness of outliers, which typically prevent us from collecting a representative set of anomalous sounds. As a consequence, most anomaly detection methods use unsupervised rather than supervised machine learning methods. Nevertheless, we will show that anomalous sound detection can be effectively framed as a supervised classification problem if the set of anomalous samples is carefully substituted with what we call proxy outliers. Candidates for proxy outliers are available in abundance as they potentially include all recordings that are neither normal nor abnormal sounds. We experiment with the machine condition monitoring data set of the 2020's DCASE Challenge and find proxy outliers with matching recording conditions and high similarity to the target sounds particularly informative. If no data with similar sounds and matching recording conditions is available, data sets with a larger diversity in these two dimensions are preferable. Our models based on supervised training with proxy outliers achieved rank three in Task 2 of the DCASE2020 Challenge.

Khaled Koutini, Hamid Eghbal-Zadeh, Verena Haunschmid et al.

Convolutional Neural Networks (CNNs) have been successfully used in various Music Information Retrieval (MIR) tasks, both as end-to-end models and as feature extractors for more complex systems. However, the MIR field is still dominated by the classical VGG-based CNN architecture variants, often in combination with more complex modules such as attention, and/or techniques such as pre-training on large datasets. Deeper models such as ResNet -- which surpassed VGG by a large margin in other domains -- are rarely used in MIR. One of the main reasons for this, as we will show, is the lack of generalization of deeper CNNs in the music domain. In this paper, we present a principled way to make deep architectures like ResNet competitive for music-related tasks, based on well-designed regularization strategies. In particular, we analyze the recently introduced Receptive-Field Regularization and Shake-Shake, and show that they significantly improve the generalization of deep CNNs on music-related tasks, and that the resulting deep CNNs can outperform current more complex models such as CNNs augmented with pre-training and attention. We demonstrate this on two different MIR tasks and two corresponding datasets, thus offering our deep regularized CNNs as a new baseline for these datasets, which can also be used as a feature-extracting module in future, more complex approaches.

Hamid Eghbal-zadeh, Khaled Koutini, Paul Primus et al.

Data augmentation techniques have become standard practice in deep learning, as it has been shown to greatly improve the generalisation abilities of models. These techniques rely on different ideas such as invariance-preserving transformations (e.g, expert-defined augmentation), statistical heuristics (e.g, Mixup), and learning the data distribution (e.g, GANs). However, in the adversarial settings it remains unclear under what conditions such data augmentation methods reduce or even worsen the misclassification risk. In this paper, we therefore analyse the effect of different data augmentation techniques on the adversarial risk by three measures: (a) the well-known risk under adversarial attacks, (b) a new measure of prediction-change stress based on the Laplacian operator, and (c) the influence of training examples on prediction. The results of our empirical analysis disprove the hypothesis that an improvement in the classification performance induced by a data augmentation is always accompanied by an improvement in the risk under adversarial attack. Further, our results reveal that the augmented data has more influence than the non-augmented data, on the resulting models. Taken together, our results suggest that general-purpose data augmentations that do not take into the account the characteristics of the data and the task, must be applied with care.

Paul Primus, Hamid Eghbal-zadeh, David Eitelsebner et al.

Distribution mismatches between the data seen at training and at application time remain a major challenge in all application areas of machine learning. We study this problem in the context of machine listening (Task 1b of the DCASE 2019 Challenge). We propose a novel approach to learn domain-invariant classifiers in an end-to-end fashion by enforcing equal hidden layer representations for domain-parallel samples, i.e. time-aligned recordings from different recording devices. No classification labels are needed for our domain adaptation (DA) method, which makes the data collection process cheaper.