Music is present in every known society but varies from place to place. What, if anything, is universal to music cognition? We measured a signature of mental representations of rhythm in 39 participant groups in 15 countries, spanning urban societies and Indigenous populations. Listeners reproduced random ‘seed’ rhythms; their reproductions were fed back as the stimulus (as in the game of ‘telephone’), such that their biases (the prior) could be estimated from the distribution of reproductions. Every tested group showed a sparse prior with peaks at integer-ratio rhythms. However, the importance of different integer ratios varied across groups, often reflecting local musical practices. Our results suggest a common feature of music cognition: discrete rhythm ‘categories’ at small-integer ratios. These discrete representations plausibly stabilize musical systems in the face of cultural transmission but interact with culture-specific traditions to yield the diversity that is evident when mental representations are probed across many cultures.
Deep learning models have improved cutting-edge technologies in many research areas, but their black-box structure makes it difficult to understand their inner workings and the rationale behind their predictions. This may lead to unintended effects, such as being susceptible to adversarial attacks or the reinforcement of biases. There is still a lack of research in the audio domain, despite the increasing interest in developing deep learning models that provide explanations of their decisions. To reduce this gap, we propose a novel interpretable deep learning model for automatic sound classification, which explains its predictions based on the similarity of the input to a set of learned prototypes in a latent space. We leverage domain knowledge by designing a frequency-dependent similarity measure and by considering different time-frequency resolutions in the feature space. The proposed model achieves results that are comparable to that of the state-of-the-art methods in three different sound classification tasks involving speech, music, and environmental audio. In addition, we present two automatic methods to prune the proposed model that exploit its interpretability. Our system is open source and it is accompanied by a web application for the manual editing of the model, which allows for a human-in-the-loop debugging approach.
Both music and language are found in all known human societies, yet no studies have compared similarities and differences between song, speech, and instrumental music on a global scale. In this Registered Report, we analyzed two global datasets: (i) 300 annotated audio recordings representing matched sets of traditional songs, recited lyrics, conversational speech, and instrumental melodies from our 75 coauthors speaking 55 languages; and (ii) 418 previously published adult-directed song and speech recordings from 209 individuals speaking 16 languages. Of our six preregistered predictions, five were strongly supported: Relative to speech, songs use (i) higher pitch, (ii) slower temporal rate, and (iii) more stable pitches, while both songs and speech used similar (iv) pitch interval size and (v) timbral brightness. Exploratory analyses suggest that features vary along a “musi-linguistic” continuum when including instrumental melodies and recited lyrics. Our study provides strong empirical evidence of cross-cultural regularities in music and speech.
Training state-of-the-art beat tracking models usually requires large amounts of annotated data. It is widely known that data annotation is a time-consuming process and generally involves expert knowledge in the context of MIR. This can be particularly challenging if we consider culture-specific datasets. Previous research has shown that, under certain homogeneity conditions, it is possible to obtain good tracking results with these models using few training datapoints. However, this shifts the problem to that of the selection of these data. In this paper, we propose a methodology for selectively annotating meaningful samples from a dataset with the objective of training a beat tracker. We extract a rhythmic feature from each track and apply selection methods in the feature space limited by a budget of samples to be annotated. We then train a TCN-based state-of-the-art model using the selected data. The trained model is shown to perform well on the remainder of the dataset when compared to random selection. We hope that our study will alleviate the annotation process of culture-specific datasets and ultimately help build a more culturally diverse perspective in the field of Music Information Retrieval.