Bob L. Sturm

Simon Colton, Maria Teresa Llano, Rose Hepworth et al.

During 2015 and early 2016, the cultural application of Computational Creativity research and practice took a big leap forward, with a project where multiple computational systems were used to provide advice and material for a new musical theatre production. Billed as the world's first 'computer musical... conceived by computer and substantially crafted by computer', Beyond The Fence was staged in the Arts Theatre in London's West End during February and March of 2016. Various computational approaches to analytical and generative sub-projects were used to bring about the musical, and these efforts were recorded in two 1-hour documentary films made by Wingspan Productions, which were aired on SkyArts under the title Computer Says Show. We provide details here of the project conception and execution, including details of the systems which took on some of the creative responsibility in writing the musical, and the contributions they made. We also provide details of the impact of the project, including a perspective from the two (human) writers with overall control of the creative aspects the musical.

Nicolas Jonason, Bob L. T. Sturm

We present work in progress on TimbreCLIP, an audio-text cross modal embedding trained on single instrument notes. We evaluate the models with a cross-modal retrieval task on synth patches. Finally, we demonstrate the application of TimbreCLIP on two tasks: text-driven audio equalization and timbre to image generation.

Nicolas Jonason, Bob L. T. Sturm

We explore the generation of visualisations of audio latent spaces using an audio-to-image generation pipeline. We believe this can help with the interpretability of audio latent spaces. We demonstrate a variety of results on the NSynth dataset. A web demo is available.

Luca Casini, Laura Cros Vila, David Dalmazzo et al.

Online AI platforms for creating music from text prompts (AI music), such as Suno and Udio, are now being used by hundreds of thousands of users. Some AI music is appearing in advertising, and even charting, in multiple countries. How are these platforms being used? What subjects are inspiring their users? This article answers these questions for Suno and Udio using a large collection of songs generated by users of these platforms from May to October 2024. Using a combination of state-of-the-art text embedding models, dimensionality reduction and clustering methods, we analyze the prompts, tags and lyrics, and automatically annotate and display the processed data in interactive plots. Our results reveal prominent themes in lyrics, language preference, prompting strategies, as well as peculiar attempts at steering models through the use of metatags. To promote the musicological study of the developing cultural practice of AI-generated music we share our code and resources.

Bob L. Sturm, João Felipe Santos, Oded Ben-Tal et al.

We apply deep learning methods, specifically long short-term memory (LSTM) networks, to music transcription modelling and composition. We build and train LSTM networks using approximately 23,000 music transcriptions expressed with a high-level vocabulary (ABC notation), and use them to generate new transcriptions. Our practical aim is to create music transcription models useful in particular contexts of music composition. We present results from three perspectives: 1) at the population level, comparing descriptive statistics of the set of training transcriptions and generated transcriptions; 2) at the individual level, examining how a generated transcription reflects the conventions of a music practice in the training transcriptions (Celtic folk); 3) at the application level, using the system for idea generation in music composition. We make our datasets, software and sound examples open and available: \url{https://github.com/IraKorshunova/folk-rnn}.

Nicolas Jonason, Luca Casini, Bob L. T. Sturm

We present a new approach for fast and controllable generation of symbolic music based on the simplex diffusion, which is essentially a diffusion process operating on probabilities rather than the signal space. This objective has been applied in domains such as natural language processing but here we apply it to generating 4-bar multi-instrument music loops using an orderless representation. We show that our model can be steered with vocabulary priors, which affords a considerable level control over the music generation process, for instance, infilling in time and pitch and choice of instrumentation -- all without task-specific model adaptation or applying extrinsic control.

Bob L. T. Sturm

I reflect on my experience creating two music albums centered on state-of-the-art prompt-based AI music generation platforms. The first album explicitly poses the question: What happens when I collide my junk mail with these platforms? The second album is a direct response to the first, and toys with the inability of state-of-the-art prompt-based AI music generation platforms to generate music that is not ``practiced'', ``polished'', and ``produced''. I seed a large language model (LLM) with information about these albums and have it interview me, which results in the exploration of several deeper questions: To what extent am I the author? Where am I in the resulting music? How is my musical identity changing as I am faced with machines that are in some ways far more talented than I? What new musical spaces does my work open, for me or anyone/thing else? I conclude by reflecting on my reflections, as well as LLM-mediated self-reflection as method.

Nicolas Jonason, Bob L. T. Sturm

This document presents some early explorations of applying Softly Masked Language Modelling (SMLM) to symbolic music generation. SMLM can be seen as a generalisation of masked language modelling (MLM), where instead of each element of the input set being either known or unknown, each element can be known, unknown or partly known. We demonstrate some results of applying SMLM to constrained symbolic music generation using a transformer encoder architecture. Several audio examples are available at https://erl-j.github.io/smlm-web-supplement/

Bhusan Chettri, Emmanouil Benetos, Bob L. T. Sturm

The Automatic Speaker Verification Spoofing and Countermeasures Challenges motivate research in protecting speech biometric systems against a variety of different access attacks. The 2017 edition focused on replay spoofing attacks, and involved participants building and training systems on a provided dataset (ASVspoof 2017). More than 60 research papers have so far been published with this dataset, but none have sought to answer why countermeasures appear successful in detecting spoofing attacks. This article shows how artefacts inherent to the dataset may be contributing to the apparent success of published systems. We first inspect the ASVspoof 2017 dataset and summarize various artefacts present in the dataset. Second, we demonstrate how countermeasure models can exploit these artefacts to appear successful in this dataset. Third, for reliable and robust performance estimates on this dataset we propose discarding nonspeech segments and silence before and after the speech utterance during training and inference. We create speech start and endpoint annotations in the dataset and demonstrate how using them helps countermeasure models become less vulnerable from being manipulated using artefacts found in the dataset. Finally, we provide several new benchmark results for both frame-level and utterance-level models that can serve as new baselines on this dataset.

Saumitra Mishra, Emmanouil Benetos, Bob L. Sturm et al.

One way to analyse the behaviour of machine learning models is through local explanations that highlight input features that maximally influence model predictions. Sensitivity analysis, which involves analysing the effect of input perturbations on model predictions, is one of the methods to generate local explanations. Meaningful input perturbations are essential for generating reliable explanations, but there exists limited work on what such perturbations are and how to perform them. This work investigates these questions in the context of machine listening models that analyse audio. Specifically, we use a state-of-the-art deep singing voice detection (SVD) model to analyse whether explanations from SoundLIME (a local explanation method) are sensitive to how the method perturbs model inputs. The results demonstrate that SoundLIME explanations are sensitive to the content in the occluded input regions. We further propose and demonstrate a novel method for quantitatively identifying suitable content type(s) for reliably occluding inputs of machine listening models. The results for the SVD model suggest that the average magnitude of input mel-spectrogram bins is the most suitable content type for temporal explanations.

Saumitra Mishra, Daniel Stoller, Emmanouil Benetos et al.

One way to interpret trained deep neural networks (DNNs) is by inspecting characteristics that neurons in the model respond to, such as by iteratively optimising the model input (e.g., an image) to maximally activate specific neurons. However, this requires a careful selection of hyper-parameters to generate interpretable examples for each neuron of interest, and current methods rely on a manual, qualitative evaluation of each setting, which is prohibitively slow. We introduce a new metric that uses Fréchet Inception Distance (FID) to encourage similarity between model activations for real and generated data. This provides an efficient way to evaluate a set of generated examples for each setting of hyper-parameters. We also propose a novel GAN-based method for generating explanations that enables an efficient search through the input space and imposes a strong prior favouring realistic outputs. We apply our approach to a classification model trained to predict whether a music audio recording contains singing voice. Our results suggest that this proposed metric successfully selects hyper-parameters leading to interpretable examples, avoiding the need for manual evaluation. Moreover, we see that examples synthesised to maximise or minimise the predicted probability of singing voice presence exhibit vocal or non-vocal characteristics, respectively, suggesting that our approach is able to generate suitable explanations for understanding concepts learned by a neural network.

Bhusan Chettri, Daniel Stoller, Veronica Morfi et al.

Detecting spoofing attempts of automatic speaker verification (ASV) systems is challenging, especially when using only one modeling approach. For robustness, we use both deep neural networks and traditional machine learning models and combine them as ensemble models through logistic regression. They are trained to detect logical access (LA) and physical access (PA) attacks on the dataset released as part of the ASV Spoofing and Countermeasures Challenge 2019. We propose dataset partitions that ensure different attack types are present during training and validation to improve system robustness. Our ensemble model outperforms all our single models and the baselines from the challenge for both attack types. We investigate why some models on the PA dataset strongly outperform others and find that spoofed recordings in the dataset tend to have longer silences at the end than genuine ones. By removing them, the PA task becomes much more challenging, with the tandem detection cost function (t-DCF) of our best single model rising from 0.1672 to 0.5018 and equal error rate (EER) increasing from 5.98% to 19.8% on the development set.

Bhusan Chettri, Saumitra Mishra, Bob L. Sturm et al.

The second Automatic Speaker Verification Spoofing and Countermeasures challenge (ASVspoof 2017) focused on "replay attack" detection. The best deep-learning systems to compete in ASVspoof 2017 used Convolutional Neural Networks (CNNs) as a feature extractor. In this paper, we study their performance in an end-to-end setting. We find that these architectures show poor generalization in the evaluation dataset, but find a compact architecture that shows good generalization on the development data. We demonstrate that for this dataset it is not easy to obtain a similar level of generalization on both the development and evaluation data. This leads to a variety of open questions about what the differences are in the data; why these are more evident in an end-to-end setting; and how these issues can be overcome by increasing the training data.

Bob L. Sturm

Building systems that possess the sensitivity and intelligence to identify and describe high-level attributes in music audio signals continues to be an elusive goal, but one that surely has broad and deep implications for a wide variety of applications. Hundreds of papers have so far been published toward this goal, and great progress appears to have been made. Some systems produce remarkable accuracies at recognising high-level semantic concepts, such as music style, genre and mood. However, it might be that these numbers do not mean what they seem. In this paper, we take a state-of-the-art music content analysis system and investigate what causes it to achieve exceptionally high performance in a benchmark music audio dataset. We dissect the system to understand its operation, determine its sensitivities and limitations, and predict the kinds of knowledge it could and could not possess about music. We perform a series of experiments to illuminate what the system has actually learned to do, and to what extent it is performing the intended music listening task. Our results demonstrate how the initial manifestation of music intelligence in this state-of-the-art can be deceptive. Our work provides constructive directions toward developing music content analysis systems that can address the music information and creation needs of real-world users.

Corey Kereliuk, Bob L. Sturm, Jan Larsen

An adversary is essentially an algorithm intent on making a classification system perform in some particular way given an input, e.g., increase the probability of a false negative. Recent work builds adversaries for deep learning systems applied to image object recognition, which exploits the parameters of the system to find the minimal perturbation of the input image such that the network misclassifies it with high confidence. We adapt this approach to construct and deploy an adversary of deep learning systems applied to music content analysis. In our case, however, the input to the systems is magnitude spectral frames, which requires special care in order to produce valid input audio signals from network-derived perturbations. For two different train-test partitionings of two benchmark datasets, and two different deep architectures, we find that this adversary is very effective in defeating the resulting systems. We find the convolutional networks are more robust, however, compared with systems based on a majority vote over individually classified audio frames. Furthermore, we integrate the adversary into the training of new deep systems, but do not find that this improves their resilience against the same adversary.

Fabien Gouyon, Bob L. Sturm, Joao Lobato Oliveira et al.

Music autotagging, an established problem in Music Information Retrieval, aims to alleviate the human cost required to manually annotate collections of recorded music with textual labels by automating the process. Many autotagging systems have been proposed and evaluated by procedures and datasets that are now standard (used in MIREX, for instance). Very little work, however, has been dedicated to determine what these evaluations really mean about an autotagging system, or the comparison of two systems, for the problem of annotating music in the real world. In this article, we are concerned with explaining the figure of merit of an autotagging system evaluated with a standard approach. Specifically, does the figure of merit, or a comparison of figures of merit, warrant a conclusion about how well autotagging systems have learned to describe music with a specific vocabulary? The main contributions of this paper are a formalization of the notion of validity in autotagging evaluation, and a method to test it in general. We demonstrate the practical use of our method in experiments with three specific state-of-the-art autotagging systems --all of which are reproducible using the linked code and data. Our experiments show for these specific systems in a simple and objective two-class task that the standard evaluation approach does not provide valid indicators of their performance.

Bob L. Sturm

The GTZAN dataset appears in at least 100 published works, and is the most-used public dataset for evaluation in machine listening research for music genre recognition (MGR). Our recent work, however, shows GTZAN has several faults (repetitions, mislabelings, and distortions), which challenge the interpretability of any result derived using it. In this article, we disprove the claims that all MGR systems are affected in the same ways by these faults, and that the performances of MGR systems in GTZAN are still meaningfully comparable since they all face the same faults. We identify and analyze the contents of GTZAN, and provide a catalog of its faults. We review how GTZAN has been used in MGR research, and find few indications that its faults have been known and considered. Finally, we rigorously study the effects of its faults on evaluating five different MGR systems. The lesson is not to banish GTZAN, but to use it with consideration of its contents.