Automatic drum transcription (ADT) is traditionally formulated as a discriminative task to predict drum events from audio spectrograms. In this work, we redefine ADT as a conditional generative task and introduce Noise-to-Notes (N2N), a framework leveraging diffusion modeling to transform audio-conditioned Gaussian noise into drum events with associated velocities. This generative diffusion approach offers distinct advantages, including a flexible speed-accuracy trade-off and strong inpainting capabilities. However, the generation of binary onset and continuous velocity values presents a challenge for diffusion models, and to overcome this, we introduce an Annealed Pseudo-Huber loss to facilitate effective joint optimization. Finally, to augment low-level spectrogram features, we propose incorporating features extracted from music foundation models (MFMs), which capture high-level semantic information and enhance robustness to out-of-domain drum audio. Experimental results demonstrate that including MFM features significantly improves robustness and N2N establishes a new state-of-the-art performance across multiple ADT benchmarks.

Automatic drum transcription is a critical tool in Music Information Retrieval for extracting and analyzing the rhythm of a music track, but it is limited by the size of the datasets available for training. A popular method used to increase the amount of data is by generating them synthetically from music scores rendered with virtual instruments. This method can produce a virtually infinite quantity of tracks, but empirical evidence shows that models trained on previously created synthetic datasets do not transfer well to real tracks. In this work, besides increasing the amount of data, we identify and evaluate three more strategies that practitioners can use to improve the realism of the generated data and, thus, narrow the synthetic-to-real transfer gap. To explore their efficacy, we used them to build a new synthetic dataset and then we measured how the performance of a model scales and, specifically, at what value it will stagnate when increasing the number of training tracks for different datasets. By doing this, we were able to prove that the aforementioned strategies contribute to make our dataset the one with the most realistic data distribution and the lowest synthetic-to-real transfer gap among the synthetic datasets we evaluated. We conclude by highlighting the limits of training with infinite data in drum transcription and we show how they can be overcome.

In the past, the field of drum source separation faced significant challenges due to limited data availability, hindering the adoption of cutting-edge deep learning methods that have found success in other related audio applications. In this manuscript, we introduce StemGMD, a large-scale audio dataset of isolated single-instrument drum stems. Each audio clip is synthesized from MIDI recordings of expressive drums performances using ten real-sounding acoustic drum kits. Totaling 1224 hours, StemGMD is the largest audio dataset of drums to date and the first to comprise isolated audio clips for every instrument in a canonical nine-piece drum kit. We leverage StemGMD to develop LarsNet, a novel deep drum source separation model. Through a bank of dedicated U-Nets, LarsNet can separate five stems from a stereo drum mixture faster than real-time and is shown to significantly outperform state-of-the-art nonnegative spectro-temporal factorization methods.

Real-time music information retrieval (RT-MIR) has much potential to augment the capabilities of traditional acoustic instruments. We develop RT-MIR techniques aimed at augmenting percussive fingerstyle, which blends acoustic guitar playing with guitar body percussion. We formulate several design objectives for RT-MIR systems for augmented instrument performance: (i) causal constraint, (ii) perceptually negligible action-to-sound latency, (iii) control intimacy support, (iv) synthesis control support. We present and evaluate real-time guitar body percussion recognition and embedding learning techniques based on convolutional neural networks (CNNs) and CNNs jointly trained with variational autoencoders (VAEs). We introduce a taxonomy of guitar body percussion based on hand part and location. We follow a cross-dataset evaluation approach by collecting three datasets labelled according to the taxonomy. The embedding quality of the models is assessed using KL-Divergence across distributions corresponding to different taxonomic classes. Results indicate that the networks are strong classifiers especially in a simplified 2-class recognition task, and the VAEs yield improved class separation compared to CNNs as evidenced by increased KL-Divergence across distributions. We argue that the VAE embedding quality could support control intimacy and rich interaction when the latent space's parameters are used to control an external synthesis engine. Further design challenges around generalisation to different datasets have been identified.

We introduce DrummerNet, a drum transcription system that is trained in an unsupervised manner. DrummerNet does not require any ground-truth transcription and, with the data-scalability of deep neural networks, learns from a large unlabeled dataset. In DrummerNet, the target drum signal is first passed to a (trainable) transcriber, then reconstructed in a (fixed) synthesizer according to the transcription estimate. By training the system to minimize the distance between the input and the output audio signals, the transcriber learns to transcribe without ground truth transcription. Our experiment shows that DrummerNet performs favorably compared to many other recent drum transcription systems, both supervised and unsupervised.

We propose a novel model for temporal detection and localization which allows the training of deep neural networks using only counts of event occurrences as training labels. This powerful weakly-supervised framework alleviates the burden of the imprecise and time-consuming process of annotating event locations in temporal data. Unlike existing methods, in which localization is explicitly achieved by design, our model learns localization implicitly as a byproduct of learning to count instances. This unique feature is a direct consequence of the model's theoretical properties. We validate the effectiveness of our approach in a number of experiments (drum hit and piano onset detection in audio, digit detection in images) and demonstrate performance comparable to that of fully-supervised state-of-the-art methods, despite much weaker training requirements.