Umeå universitets logga

The automatic identification of cue points is a central task in applications as diverse as music thumbnailing, generation of mash ups, and DJ mixing. Our focus lies in electronic dance music and in a specific kind of cue point, the “switch point,” that makes it possible to automatically construct transitions between tracks, mimicking what professional DJs do. We present two approaches for the detection of switch points. One embodies a few general rules we established from interviews with professional DJs, the other models a manually annotated dataset that we curated. Both approaches are based on feature extraction and novelty analysis. From an evaluation conducted on previously unknown tracks, we found that about 90 percent of the points generated can be reliably used in the context of a DJ mix.

The state-of-the-art methods for drum transcription in the presence of melodic instruments (DTM) are machine learning models trained in a supervised manner, which means that they rely on labeled datasets. The problem is that the available public datasets are limited either in size or in realism, and are thus suboptimal for training purposes. Indeed, the best results are currently obtained via a rather convoluted multi-step training process that involves both real and synthetic datasets. To address this issue, starting from the observation that the communities of rhythm games players provide a large amount of annotated data, we curated a new dataset of crowdsourced drum transcriptions. This dataset contains real-world music, is manually annotated, and is about two orders of magnitude larger than any other non-synthetic dataset, making it a prime candidate for training purposes. However, due to crowdsourcing, the initial annotations contain mistakes. We discuss how the quality of the dataset can be improved by automatically correcting different types of mistakes. When used to train a popular DTM model, the dataset yields a performance that matches that of the state-of-the-art for DTM, thus demonstrating the quality of the annotations.

Within the broad problem known as automatic music transcription, we considered the specific task of automatic drum transcription (ADT). This is a complex task that has recently shown significant advances thanks to deep learning (DL) techniques. Most notably, massive amounts of labeled data obtained from crowds of annotators have made it possible to implement large-scale supervised learning architectures for ADT. In this study, we explored the untapped potential of these new datasets by addressing three key points: First, we reviewed recent trends in DL architectures and focused on two techniques, self-attention mechanisms and tatum-synchronous convolutions. Then, to mitigate the noise and bias that are inherent in crowdsourced data, we extended the training data with additional annotations. Finally, to quantify the potential of the data, we compared many training scenarios by combining up to six different datasets, including zero-shot evaluations. Our findings revealed that crowdsourced datasets outperform previously utilized datasets, and regardless of the DL architecture employed, they are sufficient in size and quality to train accurate models. By fully exploiting this data source, our models produced high-quality drum transcriptions, achieving state-of-the-art results. Thanks to this accuracy, our work can be more successfully used by musicians (e.g., to learn new musical pieces by reading, or to convert their performances to MIDI) and researchers in music information retrieval (e.g., to retrieve information from the notes instead of audio, such as the rhythm or structure of a piece).

The importance of automatic drum transcription lies in the potential to extract useful information from a musical track; however, the low reliability of the models for this task represents a limiting factor. Indeed, even though in the recent literature the quality of the generated transcription has improved thanks to the curation of large training datasets via crowdsourcing, there is still a large margin of improvement for this task to be considered solved. Aiming to steer the development of future models, we identify the most common errors from training and testing on the aforementioned crowdsourced datasets. We perform this study in three steps: First, we detail the quality of the transcription for each class of interest; second, we employ a new metric and a pseudo confusion matrix to quantify different mistakes in the estimations; last, we compute the agreement between different annotators of the same track to estimate the accuracy of the ground-truth. Our findings are twofold: On the one hand, we observe that the previously reported issue that less represented instruments (e.g., toms) are less reliably transcribed is mostly solved now. On the other hand, cymbal instruments have unprecedented relative low performance. We provide intuitive explanations as to why cymbal instruments are difficult to transcribe and we identify that they represent the main source of disagreement among annotators.