Download Empirical Results for Adjusting Truncated Backpropagation Through Time While Training Neural Audio Effects
This paper investigates the optimization of Truncated Backpropagation Through Time (TBPTT) for training neural networks in digital audio effect modeling, with a focus on dynamic range compression. The study evaluates key TBPTT hyperparameters – sequence number, batch size, and sequence length – and their influence on model performance. Using a convolutional-recurrent architecture, we conduct extensive experiments across datasets with and without conditioning by user controls. Results demonstrate that carefully tuning these parameters enhances model accuracy and training stability, while also reducing computational demands. Objective evaluations confirm improved performance with optimized settings, while subjective listening tests indicate that the revised TBPTT configuration maintains high perceptual quality.
Download Wave Digital Filter Adaptors for Arbitrary Topologies and Multiport Linear Elements
We present a Modified-Nodal-Analysis-derived method for developing Wave Digital Filter (WDF) adaptors corresponding to complicated (non-series/parallel) topologies that may include multiport linear elements (e.g. controlled sources and transformers). A second method resolves noncomputable (non-tree-like) arrangements of series/parallel adaptors. As with the familiar 3-port series and parallel adaptors, one port of each derived adaptor may be rendered reflection-free, making it acceptable for inclusion in a standard WDF tree. With these techniques, the class of acceptable reference circuits for WDF modeling is greatly expanded. This is demonstrated by case studies on circuits which were previously intractable with WDF methods: the Bassman tone stack and Tube Screamer tone/volume stage.
Download Benchmarking Integrated GPU Acceleration of Real-Time Neural Audio Inference on Snapdragon
This paper investigates whether integrated GPUs on Qualcomm Snapdragon SoCs can accelerate streaming inference of neural audio models. Five models spanning three orders of magnitude in parameter count are benchmarked across three inference approaches (best available CPU, QNN CPU, and QNN GPU). Results reveal when GPU acceleration offers meaningful gains, when per-call overhead negates benefits, and how model size and architecture determine GPU suitability.
Download Application of non-negative matrix factorization to signal-adaptive audio effects
This paper proposes novel audio effects based on manipulating an audio signal in a representation domain provided by non-negative matrix factorization (NMF). Critical-band magnitude spectrograms Y of sounds are first factorized into a product of two lower-rank matrices so that Y ≈ BG. The parameter matrices B and G are then processed in order to achieve the desired effect. Three classes of effects were investigated: 1) dynamic range compression (or expansion) of the component spectra or gains, 2) effects based on rank-ordering the components (colums of B and the corresponding rows of G) according to acoustic features extracted from them, and then weighting each component according to its rank, and 3) distortion effects based on controlling the amount of components (and thus the reconstruction error) in the above linear approximation. The subjective quality of the effects was assessed in a listening test.
Download An Efficient Pitch-Tracking Algorithm Using A Combination Of Fourier Transforms
In this paper we present a technique for detecting the pitch of sound using a series of two forward Fourier transforms. We use an enhanced version of the Fourier transform for a better accuracy, as well as a tracking strategy among pitch candidates for an increased robustness. This efficient technique allows us to precisely find out the pitches of harmonic sounds such as the voice or classic musical instruments, but also of more complex sounds like rippled noises.
Download Analysis of transient musical sounds by auto-regressive modeling
This paper gives an example of an auto-regressive (AR) spectral analysis on transient musical sounds. The attack part of many musical sounds is mostly too short to be analysed by a short-time Fourier analysis, whereas this short period of time is long enough for several AR-methods. The AR-spectra obtained from short segments of signals with attack transients have a sufficiently high frequency resolution. These spectra contain more information about the evolution of a sound than a fast Fourier transform made over a small amount of samples.
Download Sound Source Separation: Azimuth Discrimination and Resynthesis
In this paper we present a novel sound source separation algorithm which requires no prior knowledge, no learning, assisted or otherwise, and performs the task of separation based purely on azimuth discrimination within the stereo field. The algorithm exploits the use of the pan pot as a means to achieve image localisation within stereophonic recordings. As such, only an interaural intensity difference exists between left and right channels for a single source. We use gain scaling and phase cancellation techniques to expose frequency dependent nulls across the azimuth domain, from which source separation and resynthesis is carried out. We present results obtained from real recordings, and show that for musical recordings, the algorithm improves upon the output quality of current source separation schemes.
Download Acoustical Simulations of the Human Vocal Tract Using the 1D and 2D Digital Waveguide Software Model
This paper details software under development that uses the digital waveguide physical model to represent the sound creation mechanism and environment associated with the production of speech, specifically the human vocal tract. Focus is directed towards a comparison between the existing 1D waveguide method, on which several studies have already been conducted, and the developing 2D waveguide mesh method. The construction of the two models and the application of the tract geometry is examined, in addition, the inclusion of dynamic articulatory variations to increase the ability of such systems to create natural sounding speech is discussed. Results obtained from each suggest that the 2D model is capable of producing similarly accurate vowel spectra to that already accomplished with the 1D version, although speech-like sounds created with the 2D mesh appear to exhibit greater realism.
Download Implementation of Arbitrary Linear Sound Synthesis Algorithms by Digital Wave Guide Structures
The Digital Wave Guide (DWG) method is one of the most popular techniques for digital sound synthesis via physical modeling. Due to the inherent solution of the wave equation by the structure of the DWG method, it provides a highly efficient algorithm for typical physical modeling problems. In this paper it will be shown, that it is possible to use this efficient structure for any existing linear sound synthesis algorithm. By a consequent description of discrete implementations with State Space Structures (SSSs), suitable linear state space transformations can be used to achieve the typical DWG structure from any given system. The proposed approach is demonstrated with two case studies, where a modal solution achieved with the Functional Transformation Method (FTM) is transformed to a DWG implementation. In the first example the solution of the lossless wave equation is transformed to a DWG structure, yielding an arbitrary size fractional delay filter. In another example a more elaborated model with dispersion and damping terms is transformed, resulting in a DWG model with parameter morphing features.
Download Separation of overlapping impulsive sounds by bandwise noise interpolation
The task of extracting harmonic content of multiple pitched sources from a mono audio mix has been investigated on several occasions [1, 2, 3, 4]. However, most pitched notes contain an inharmonic component, which is an important perceptual attribute of the sound. This content is usually not dealt with during separation. It would also be interesting in its own right to develop separation techniques for extracting percussive sounds for polyphonic mixes. This paper describes an attempt at separating overlapping impulsive content of multiple sources from a mono mix. The method uses an interpolation within individual frequency bands of the decaying noise envelope of each source across overlapping sections with other sources. Three analysis methods determining the distribution of these bands were tested: the DFT followed by processing in Bark bands, the discrete wavelet transform (DWT), and the dyadic wavelet packet transform (DWPT).