Download Unsupervised Taxonomy of Sound Effects Sound effect libraries are commonly used by sound designers in a range of industries. Taxonomies exist for the classification of sounds into groups based on subjective similarity, sound source or common environmental context. However, these taxonomies are not standardised, and no taxonomy based purely on the sonic properties of audio exists. We present a method using feature selection, unsupervised learning and hierarchical clustering to develop an unsupervised taxonomy of sound effects based entirely on the sonic properties of the audio within a sound effect library. The unsupervised taxonomy is then related back to the perceived meaning of the relevant audio features.
Download Latent Force Models for Sound: Learning Modal Synthesis Parameters and Excitation Functions from Audio Recordings Latent force models are a Bayesian learning technique that combine physical knowledge with dimensionality reduction — sets of coupled differential equations are modelled via shared dependence on a low-dimensional latent space. Analogously, modal sound synthesis is a technique that links physical knowledge about the vibration of objects to acoustic phenomena that can be observed in data. We apply latent force modelling to sinusoidal models of audio recordings, simultaneously inferring modal synthesis parameters (stiffness and damping) and the excitation or contact force required to reproduce the behaviour of the observed vibrational modes. Exposing this latent excitation function to the user constitutes a controllable synthesis method that runs in real time and enables sound morphing through interpolation of learnt parameters.
Download End-to-end equalization with convolutional neural networks 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 Modelling Experts’ Decisions on Assigning Narrative Importances of Objects in a Radio Drama Mix There is an increasing number of consumers of broadcast audio who suffer from a degree of hearing impairment. One of the methods developed for tackling this issue consists of creating customizable object-based audio mixes where users can attenuate parts of the mix using a simple complexity parameter. The method relies on the mixing engineer classifying audio objects in the mix according to their narrative importance. This paper focuses on automating this process. Individual tracks are classified based on their music, speech, or sound effect content. Then the decisions for assigning narrative importance to each segment of a radio drama mix are modelled using mixture distributions. Finally, the learned decisions and resultant mixes are evaluated using the Short Term Objective Intelligibility, with reference to the narrative importance selections made by the original producer. This approach has applications for providing customizable mixes for legacy content, or automatically generated media content where the engineer is not able to intervene.
Download Optimization techniques for a physical model of human vocalisation We present a non-supervised approach to optimize and evaluate the synthesis of non-speech audio effects from a speech production model. We use the Pink Trombone synthesizer as a case study of a simplified production model of the vocal tract to target nonspeech human audio signals –yawnings. We selected and optimized the control parameters of the synthesizer to minimize the difference between real and generated audio. We validated the most common optimization techniques reported in the literature and a specifically designed neural network. We evaluated several popular quality metrics as error functions. These include both objective quality metrics and subjective-equivalent metrics. We compared the results in terms of total error and computational demand. Results show that genetic and swarm optimizers outperform least squares algorithms at the cost of executing slower and that specific combinations of optimizers and audio representations offer significantly different results. The proposed methodology could be used in benchmarking other physical models and audio types.
Download Automatic subgrouping of multitrack audio Subgrouping is a mixing technique where the outputs of a subset of audio tracks in a multitrack are summed to a single audio bus. This is done so that the mix engineer can apply signal processing to an entire subgroup, speed up the mix work flow and manipulate a number of audio tracks at once. In this work, we investigate which audio features from a set of 159 can be used to automatically subgroup multitrack audio. We determine a subset of audio features from the original 159 audio features to use for automatic subgrouping, by performing feature selection using a Random Forest classifier on a dataset of 54 individual multitracks. We show that by using agglomerative clustering on 5 test multitracks, the entire set of audio features incorrectly clusters 35.08% of the audio tracks, while the subset of audio features incorrectly clusters only 7.89% of the audio tracks. Furthermore, we also show that using the entire set of audio features, ten incorrect subgroups are created. However, when using the subset of audio features, only five incorrect subgroups are created. This indicates that our reduced set of audio features provides a significant increase in classification accuracy for the creation of subgroups automatically.
Download The Mix Evaluation Dataset Research on perception of music production practices is mainly concerned with the emulation of sound engineering tasks through lab-based experiments and custom software, sometimes with unskilled subjects. This can improve the level of control, but the validity, transferability, and relevance of the results may suffer from this artificial context. This paper presents a dataset consisting of mixes gathered in a real-life, ecologically valid setting, and perceptual evaluation thereof, which can be used to expand knowledge on the mixing process. With 180 mixes including parameter settings, close to 5000 preference ratings and free-form descriptions, and a diverse range of contributors from five different countries, the data offers many opportunities for music production analysis, some of which are explored here. In particular, more experienced subjects were found to be more negative and more specific in their assessments of mixes, and to increasingly agree with each other.
Download Vocal Tract Area Estimation by Gradient Descent Articulatory features can provide interpretable and flexible controls for the synthesis of human vocalizations by allowing the user to directly modify parameters like vocal strain or lip position. To make this manipulation through resynthesis possible, we need to estimate the features that result in a desired vocalization directly from audio recordings. In this work, we propose a white-box optimization technique for estimating glottal source parameters and vocal tract shapes from audio recordings of human vowels. The approach is based on inverse filtering and optimizing the frequency response of a waveguide model of the vocal tract with gradient descent, propagating error gradients through the mapping of articulatory features to the vocal tract area function. We apply this method to the task of matching the sound of the Pink Trombone, an interactive articulatory synthesizer, to a given vocalization. We find that our method accurately recovers control functions for audio generated by the Pink Trombone itself. We then compare our technique against evolutionary optimization algorithms and a neural network trained to predict control parameters from audio. A subjective evaluation finds that our approach outperforms these black-box optimization baselines on the task of reproducing human vocalizations.
Download An Evaluation of Audio Feature Extraction Toolboxes Audio feature extraction underpins a massive proportion of audio processing, music information retrieval, audio effect design and audio synthesis. Design, analysis, synthesis and evaluation often rely on audio features, but there are a large and diverse range of feature extraction tools presented to the community. An evaluation of existing audio feature extraction libraries was undertaken. Ten libraries and toolboxes were evaluated with the Cranfield Model for evaluation of information retrieval systems, reviewing the coverage, effort, presentation and time lag of a system. Comparisons are undertaken of these tools and example use cases are presented as to when toolboxes are most suitable. This paper allows a software engineer or researcher to quickly and easily select a suitable audio feature extraction toolbox.