Paper Archive

Structured pruning is a technique for reducing the computational load and memory footprint of neural networks by removing structured subsets of parameters according to a predefined schedule or ranking criterion. This paper investigates the application of structured pruning to real-time neural network audio effects, focusing on both feedforward networks and recurrent architectures. We evaluate multiple pruning strategies at inference time, without retraining, and analyze their effects on model performance. To quantify the trade-off between parameter count and audio fidelity, we construct a theoretical model of the approximation error as a function of network architecture and pruning level. The resulting bounds establish a principled relationship between pruninginduced sparsity and functional error, enabling informed deployment of neural audio effects in constrained real-time environments.

Download

This paper proposes to speed up the design of a third-order graphic equalizer by training a neural network to imitate its gain optimization. Instead of using the neural network to learn to design the graphic equalizer by optimizing its magnitude response, we present the network only with example command gains and the corresponding optimized gains, which are obtained with a previously proposed least-squares-based method. We presented this idea recently for the octave graphic equalizer with 10 band filters and extend it here to the third-octave case. Instead of a network with a single hidden layer, which we previously used, this task appears to require two hidden layers. This paper shows that good results can be reached with a neural network having 62 and 31 units in the first and the second hidden layer, respectively. After the training, the resulting network can quickly and accurately design a third-order graphic equalizer with a maximum error of 1.2 dB. The computing of the filter gains is over 350 times faster with the neural network than with the original optimization method. The method is easy to apply, and may thus lead to widespread use of accurate digital graphic equalizers.

Download

This paper proposes a grey-box neural network based approach to modelling LFO modulated time-varying effects. The neural network model receives both the unprocessed audio, as well as the LFO signal, as input. This allows complete control over the model’s LFO frequency and shape. The neural networks are trained using guitar audio, which has to be processed by the target effect and also annotated with the predicted LFO signal before training. A measurement signal based on regularly spaced chirps was used to accurately predict the LFO signal. The model architecture has been previously shown to be capable of running in real-time on a modern desktop computer, whilst using relatively little processing power. We validate our approach creating models of both a phaser and a flanger effects pedal, and theoretically it can be applied to any LFO modulated time-varying effect. In the best case, an errorto-signal ratio of 1.3% is achieved when modelling a flanger pedal, and previous work has shown that this corresponds to the model being nearly indistinguishable from the target device.

Download

The sound of magnetic recording media, such as open-reel and cassette tape recorders, is still sought after by today’s sound practitioners due to the imperfections embedded in the physics of the magnetic recording process. This paper proposes a method for digitally emulating this character using neural networks. The signal chain of the proposed system consists of three main components: the hysteretic nonlinearity and filtering jointly produced by the magnetic recording process as well as the record and playback amplifiers, the fluctuating delay originating from the tape transport, and the combined additive noise component from various electromagnetic origins. In our approach, the hysteretic nonlinear block is modeled using a recurrent neural network, while the delay trajectories and the noise component are generated using separate diffusion models, which employ U-net deep convolutional neural networks. According to the conducted objective evaluation, the proposed architecture faithfully captures the character of the magnetic tape recorder. The results of this study can be used to construct virtual replicas of vintage sound recording devices with applications in music production and audio antiquing tasks.

Download

Neural networks have become invaluable for general audio processing tasks, such as virtual analog modeling of nonlinear audio equipment. For sequence modeling tasks in particular, recurrent neural networks (RNNs) have gained widespread adoption in recent years. Their general applicability and effectiveness stems partly from their inherent nonlinearity, which makes them prone to aliasing. Recent work has explored mitigating aliasing by oversampling the network—an approach whose effectiveness is directly linked with the incurred computational costs. This work explores an alternative route by extending the antiderivative antialiasing technique to explicit, computable RNNs. Detailed applications to the Gated Recurrent Unit and Long Short-Term Memory cell are shown as case studies. The proposed technique is evaluated on multiple pre-trained guitar amplifier models, assessing its impact on the amount of aliasing and model tonality. The method is shown to reduce the models’ tendency to alias considerably across all considered sample rates while only affecting their tonality moderately, without requiring high oversampling factors. The results of this study can be used to improve sound quality in neural audio processing tasks that employ a suitable class of RNNs. Additional materials are provided in the accompanying webpage.

Download

Piano sound emulation has been an active topic of research and development for several decades. Although comprehensive physicsbased piano models have been proposed, sample-based piano emulation is still widely utilized for its computational efficiency and relative accuracy despite presenting significant memory storage requirements. This paper proposes a novel hybrid approach to sample-based piano synthesis aimed at improving the fidelity of sound emulation while reducing memory requirements for storing samples. A neural network-based model processes the sound recorded from a single example of piano key at a given velocity. The network is trained to learn the nonlinear relationship between the various velocities at which a piano key is pressed and the corresponding sound alterations. Results show that the method achieves high accuracy using a specific neural architecture that is computationally efficient, presenting few trainable parameters, and it requires memory only for one sample for each piano key.

Download

Neural music instrument cloning is an application of deep neural networks for imitating the timbre of a particular music instrument recording with a trained neural network. One can create such clones using an approach such as DDSP [1], which has been shown to achieve good synthesis quality for several instrument types [2]. However, this approach needs about ten minutes of audio data from the instrument of interest (target recording audio). In this work, we modify the DDSP architecture and apply transfer learning techniques used in speech voice cloning [3] to significantly reduce the amount of target recording audio required. We compare various cloning approaches and architectures across durations of target recording audio, ranging from four to 256 seconds. We demonstrate editing of loudness and pitch as well as timbre transfer from only 16 seconds of target recording audio. Our code is available online1 as well as many audio examples.2

Download

Neural network models are increasingly deployed in audio plugins across a wide range of applications, including amplifier emulation, effects modeling, and synthesis. This paper evaluates widely used inference options including BNNSGraph, RTNeural, LibTorch, ONNX Runtime, and anira on model architectures commonly used in neural audio plugins. The key contribution is moving beyond isolated benchmarks to evaluate performance under realistic DAW contention, constructing mix sessions with configurable plugin loads. Results show that isolated benchmarks can be misleading, and BNNSGraph proves most robust for convolutional models on Apple Silicon.

Download

Neural networks have been applied within the Wave Digital Filter (WDF) framework as data-driven models for nonlinear multi-port circuit elements. Conventionally, these models are trained on wave variables obtained by sampling the current-voltage characteristic of the considered nonlinear element before being incorporated into the circuit WDF implementation. However, isolating multi-port elements for this process can be challenging, as their nonlinear behavior often depends on dynamic effects that emerge from interactions with the surrounding circuit. In this paper, we propose a novel approach for training neural models of nonlinear multi-port elements directly within a circuit’s Wave Digital (WD) discretetime implementation, relying solely on circuit input-output voltage measurements. Exploiting the differentiability of WD simulations, we embed the neural network into the simulation process and optimize its parameters using gradient-based methods by minimizing a loss function defined over the circuit output voltage. Experimental results demonstrate the effectiveness of the proposed approach in accurately capturing the nonlinear circuit behavior, while preserving the interpretability and modularity of WDFs.

Download

Neural networks that operate directly on time-domain signals are widely used for virtual analog (VA) modeling. A key limitation of these models is their dependence on the sampling rate used during training, which becomes implicitly encoded in the learned parameters, so that changing it generally alters the realized dynamics. Although architectural modifications to recurrent neural networks have been proposed to enable sample-rate independent operation, these approaches are inherently tailored to upsampling and do not accommodate downsampling scenarios. In this manuscript, we present a VA modeling framework based on Fourier Neural Operators (FNOs) adapted to process fixed-duration audio frames. The proposed formulation defines the learned mapping over a fixed temporal support and evaluates it on uniform grids of different densities, so that a model trained at a single sampling rate can be applied at unseen sampling resolutions. Numerical results on a nonlinear transistor circuit show that the proposed model achieves competitive accuracy in upsampling scenarios while remaining directly applicable to downsampling, unlike a sample-rate independent baseline recurrent architecture.

Download