Paper Archive

Modelling of analogue devices via deep neural networks (DNNs) has gained popularity recently, but their performance is usually measured using accuracy measures alone. This paper aims to assess the performance of DNN models of a high-gain vacuum-tube guitar amplifier using additional subjective measures, including preference and realism. Furthermore, the paper explores how the performance changes when genre-specific training data is used. In five listening tests, subjects rated models of a popular high-gain guitar amplifier, the Peavey 6505, in terms of preference, realism and perceptual accuracy. Two DNN models were used: a long short-term memory recurrent neural network (LSTM-RNN) and a WaveNet-based convolutional neural network (CNN). The LSTMRNN model was shown to be more accurate when trained with genre-specific data, to the extent that it could not be distinguished from the real amplifier in ABX tests. Despite minor perceptual inaccuracies, subjects found all models to be as realistic as the target in MUSHRA-like experiments, and there was no evidence to suggest that the real amplifier was preferred to any of the models in a mix. Finally, it was observed that a low-gain excerpt was more difficult to emulate, and was therefore useful to reveal differences between the models.

Download

Modal synthesis methods are a long-standing approach for modelling distributed musical systems. In some cases extensions are possible in order to handle geometric nonlinearities. One such case is the high-amplitude vibration of a string, where geometric nonlinear effects lead to perceptually important effects including pitch glides and a dependence of brightness on striking amplitude. A modal decomposition leads to a coupled nonlinear system of ordinary differential equations. Recent work in applied machine learning approaches (in particular neural ordinary differential equations) has been used to model lumped dynamic systems such as electronic circuits automatically from data. In this work, we examine how modal decomposition can be combined with neural ordinary differential equations for modelling distributed musical systems. The proposed model leverages the analytical solution for linear vibration of system’s modes and employs a neural network to account for nonlinear dynamic behaviour. Physical parameters of a system remain easily accessible after the training without the need for a parameter encoder in the network architecture. As an initial proof of concept, we generate synthetic data for a nonlinear transverse string and show that the model can be trained to reproduce the nonlinear dynamics of the system. Sound examples are presented.

Download

A method for musical audio synthesis using autoencoding neural networks is proposed. The autoencoder is trained to compress and reconstruct magnitude short-time Fourier transform frames. The autoencoder produces a spectrogram by activating its smallest hidden layer, and a phase response is calculated using real-time phase gradient heap integration. Taking an inverse short-time Fourier transform produces the audio signal. Our algorithm is light-weight when compared to current state-of-the-art audio-producing machine learning algorithms. We outline our design process, produce metrics, and detail an open-source Python implementation of our model.

Download

Automatic drum transcription (ADT) aims to detect drum events in polyphonic music. This task is part of the more general problem of transcribing a music signal in terms of its musical score and additionally can be very interesting for extracting high level information e.g. tempo, downbeat, measure. This article has the objective to investigate the use of Convolutional Neural Networks (CNN) in the context of ADT. Two different strategies are compared. First an approach based on a CNN based detection of drum only onsets is combined with an algorithm using Non-negative Matrix Deconvolution (NMD) for drum onset transcription. Then an approach relying entirely on CNN for the detection of individual drum instruments is described. The question of which loss function is the most adapted for this task is investigated together with the question of the optimal input structure. All algorithms are evaluated using the publicly available ENST Drum database, a widely used established reference dataset, allowing easy comparison with other algorithms. The comparison shows that the purely CNN based algorithm significantly outperforms the NMD based approach, and that the results are significantly better for the snare drum, but slightly worse for both the bass drum and the hi-hat when compared to the best results published so far and ones using also a neural network model.

Download

This work aims to implement a novel deep learning architecture to perform audio processing in the context of matched equalization. Most existing methods for automatic and matched equalization show effective performance and their goal is to find a respective transfer function given a frequency response. Nevertheless, these procedures require a prior knowledge of the type of filters to be modeled. In addition, fixed filter bank architectures are required in automatic mixing contexts. Based on end-to-end convolutional neural networks, we introduce a general purpose architecture for equalization matching. Thus, by using an end-toend learning approach, the model approximates the equalization target as a content-based transformation without directly finding the transfer function. The network learns how to process the audio directly in order to match the equalized target audio. We train the network through unsupervised and supervised learning procedures. We analyze what the model is actually learning and how the given task is accomplished. We show the model performing matched equalization for shelving, peaking, lowpass and highpass IIR and FIR equalizers.

Download

While immersive music productions have become popular in recent years, music content produced during the last decades has been predominantly mixed for stereo. This paper presents a datadriven approach to automatic binaural upmixing of stereo music. The network architecture HDemucs, previously utilized for both source separation and binauralization, is leveraged for an endto-end approach to binaural upmixing. We employ two distinct datasets, demonstrating that while custom-designed training data enhances the accuracy of spatial positioning, the use of professionally mixed music yields superior spatialization. The trained networks show a capacity to process multiple simultaneous sources individually and add valid binaural cues, effectively positioning sources with an average azimuthal error of less than 11.3 ◦ . A listening test with binaural experts shows it outperforms digital signal processing-based approaches to binauralization of stereo content in terms of spaciousness while preserving audio quality.

Download

Machine learning models have become ubiquitous in modeling analog audio devices. Expanding on this line of research, our study focuses on Voltage-Controlled Oscillators of analog synthesizers. We employ black box autoregressive artificial neural networks to model the typical analog waveshapes, including triangle, square, and sawtooth. The models can be conditioned on wave frequency and type, enabling the generation of pitch envelopes and morphing across waveshapes. We conduct evaluations on both synthetic and analog datasets to assess the accuracy of various architectural variants. The LSTM variant performed better, although lower frequency ranges present particular challenges.

Download

Recently, i-vector features have entered the field of Music Information Retrieval (MIR), exhibiting highly promising performance in important tasks such as music artist recognition or music similarity estimation. The i-vector modelling approach relies on a complex processing chain that limits by the use of engineered features such as MFCCs. The goal of the present paper is to make an important step towards a truly end-to-end modelling system inspired by the i-vector pipeline, to exploit the power of Deep Neural Networks1 (DNNs) to learn optimized feature spaces and transformations. Several authors have already tried to combine the power of DNNs with i-vector features, where DNNs were used for feature extraction, scoring or classification. In this paper, we try to use neural networks for the important step of i-vector post-processing and classification for the task of music artist recognition. Specifically, we propose a novel neural network for i-vector features with a cosine-distance loss function, optimized with stochastic gradient decent (SGD). We first show that current networks do not perform well with unprocessed i-vector features, and that post-processing methods such as Within-Class Covariance Normalization (WCCN) and Linear Discriminant Analysis (LDA) are crucially important to improve the i-vector representation. We further demonstrate that these linear projections (WCCN and LDA) can not be learned using general objective functions usually used in neural networks. We examine our network on a 50-class music artist recognition dataset using i-vectors extracted from frame-level timbre features. Our experiments suggest that using our network with fully unprocessed i-vectors, we can achieve the performance of the i-vector pipeline which uses i-vector post processing methods such as LDA and WCCN.

Download

Recent studies on classifying electric guitar effects have achieved high accuracy, particularly with deep learning techniques. However, these studies often rely on simplified datasets consisting mainly of single notes rather than realistic guitar recordings. Moreover, in the specific field of effect chain estimation, the literature tends to rely on large models, making them impractical for real-time or resource-constrained applications. In this work, we recorded realistic guitar performances using four different guitars and created three datasets by applying a chain of five effects with increasing complexity: (1) fixed order and parameters, (2) fixed order with randomly sampled parameters, and (3) random order and parameters. We also propose a novel Neural Architecture Search method aimed at discovering accurate yet compact convolutional neural network models to reduce power and memory consumption. We compared its performance to a basic random search strategy, showing that our custom Neural Architecture Search outperformed random search in identifying models that balance accuracy and complexity. We found that the number of convolutional and pooling layers becomes increasingly important as dataset complexity grows, while dense layers have less impact. Additionally, among the effects, tremolo was identified as the most challenging to classify.

Download

The noise that lavalier microphones produce when rubbing against clothing (typically referred to as rustle) can be extremely difficult to automatically remove because it is highly non-stationary and overlaps with speech in both time and frequency. Recent breakthroughs in deep neural networks have led to novel techniques for separating speech from non-stationary background noise. In this paper, we apply neural network speech separation techniques to remove rustle noise, and quantitatively compare multiple deep network architectures and input spectral resolutions. We find the best performance using bidirectional recurrent networks and spectral resolution of around 20 Hz. Furthermore, we propose an ambience preservation post-processing step to minimize potential gating artifacts during pauses in speech.

Download