Simon Dixon

Sungkyun Chang, Emmanouil Benetos, Holger Kirchhoff et al.

Multi-instrument music transcription aims to convert polyphonic music recordings into musical scores assigned to each instrument. This task is challenging for modeling as it requires simultaneously identifying multiple instruments and transcribing their pitch and precise timing, and the lack of fully annotated data adds to the training difficulties. This paper introduces YourMT3+, a suite of models for enhanced multi-instrument music transcription based on the recent language token decoding approach of MT3. We enhance its encoder by adopting a hierarchical attention transformer in the time-frequency domain and integrating a mixture of experts. To address data limitations, we introduce a new multi-channel decoding method for training with incomplete annotations and propose intra- and cross-stem augmentation for dataset mixing. Our experiments demonstrate direct vocal transcription capabilities, eliminating the need for voice separation pre-processors. Benchmarks across ten public datasets show our models' competitiveness with, or superiority to, existing transcription models. Further testing on pop music recordings highlights the limitations of current models. Fully reproducible code and datasets are available with demos at \url{https://github.com/mimbres/YourMT3}.

Yixiao Zhang, Junyan Jiang, Gus Xia et al. · bytedance

Lyric interpretations can help people understand songs and their lyrics quickly, and can also make it easier to manage, retrieve and discover songs efficiently from the growing mass of music archives. In this paper we propose BART-fusion, a novel model for generating lyric interpretations from lyrics and music audio that combines a large-scale pre-trained language model with an audio encoder. We employ a cross-modal attention module to incorporate the audio representation into the lyrics representation to help the pre-trained language model understand the song from an audio perspective, while preserving the language model's original generative performance. We also release the Song Interpretation Dataset, a new large-scale dataset for training and evaluating our model. Experimental results show that the additional audio information helps our model to understand words and music better, and to generate precise and fluent interpretations. An additional experiment on cross-modal music retrieval shows that interpretations generated by BART-fusion can also help people retrieve music more accurately than with the original BART.

Yixiao Zhang, Akira Maezawa, Gus Xia et al. · bytedance

Creating music is iterative, requiring varied methods at each stage. However, existing AI music systems fall short in orchestrating multiple subsystems for diverse needs. To address this gap, we introduce Loop Copilot, a novel system that enables users to generate and iteratively refine music through an interactive, multi-round dialogue interface. The system uses a large language model to interpret user intentions and select appropriate AI models for task execution. Each backend model is specialized for a specific task, and their outputs are aggregated to meet the user's requirements. To ensure musical coherence, essential attributes are maintained in a centralized table. We evaluate the effectiveness of the proposed system through semi-structured interviews and questionnaires, highlighting its utility not only in facilitating music creation but also its potential for broader applications.

Yinghao Ma, Anders Øland, Anton Ragni et al.

In recent years, foundation models (FMs) such as large language models (LLMs) and latent diffusion models (LDMs) have profoundly impacted diverse sectors, including music. This comprehensive review examines state-of-the-art (SOTA) pre-trained models and foundation models in music, spanning from representation learning, generative learning and multimodal learning. We first contextualise the significance of music in various industries and trace the evolution of AI in music. By delineating the modalities targeted by foundation models, we discover many of the music representations are underexplored in FM development. Then, emphasis is placed on the lack of versatility of previous methods on diverse music applications, along with the potential of FMs in music understanding, generation and medical application. By comprehensively exploring the details of the model pre-training paradigm, architectural choices, tokenisation, finetuning methodologies and controllability, we emphasise the important topics that should have been well explored, like instruction tuning and in-context learning, scaling law and emergent ability, as well as long-sequence modelling etc. A dedicated section presents insights into music agents, accompanied by a thorough analysis of datasets and evaluations essential for pre-training and downstream tasks. Finally, by underscoring the vital importance of ethical considerations, we advocate that following research on FM for music should focus more on such issues as interpretability, transparency, human responsibility, and copyright issues. The paper offers insights into future challenges and trends on FMs for music, aiming to shape the trajectory of human-AI collaboration in the music realm.

Yin-Jyun Luo, Sebastian Ewert, Simon Dixon

Disentangled sequential autoencoders (DSAEs) represent a class of probabilistic graphical models that describes an observed sequence with dynamic latent variables and a static latent variable. The former encode information at a frame rate identical to the observation, while the latter globally governs the entire sequence. This introduces an inductive bias and facilitates unsupervised disentanglement of the underlying local and global factors. In this paper, we show that the vanilla DSAE suffers from being sensitive to the choice of model architecture and capacity of the dynamic latent variables, and is prone to collapse the static latent variable. As a countermeasure, we propose TS-DSAE, a two-stage training framework that first learns sequence-level prior distributions, which are subsequently employed to regularise the model and facilitate auxiliary objectives to promote disentanglement. The proposed framework is fully unsupervised and robust against the global factor collapse problem across a wide range of model configurations. It also avoids typical solutions such as adversarial training which usually involves laborious parameter tuning, and domain-specific data augmentation. We conduct quantitative and qualitative evaluations to demonstrate its robustness in terms of disentanglement on both artificial and real-world music audio datasets.

Ruchit Agrawal, Daniel Wolff, Simon Dixon

Performance-score synchronization is an integral task in signal processing, which entails generating an accurate mapping between an audio recording of a performance and the corresponding musical score. Traditional synchronization methods compute alignment using knowledge-driven and stochastic approaches, and are typically unable to generalize well to different domains and modalities. We present a novel data-driven method for structure-aware performance-score synchronization. We propose a convolutional-attentional architecture trained with a custom loss based on time-series divergence. We conduct experiments for the audio-to-MIDI and audio-to-image alignment tasks pertained to different score modalities. We validate the effectiveness of our method via ablation studies and comparisons with state-of-the-art alignment approaches. We demonstrate that our approach outperforms previous synchronization methods for a variety of test settings across score modalities and acoustic conditions. Our method is also robust to structural differences between the performance and score sequences, which is a common limitation of standard alignment approaches.

Brendan O'Connor, Simon Dixon

Previous research has shown that established techniques for spoken voice conversion (VC) do not perform as well when applied to singing voice conversion (SVC). We propose an alternative loss component in a loss function that is otherwise well-established among VC tasks, which has been shown to improve our model's SVC performance. We first trained a singer identity embedding (SIE) network on mel-spectrograms of singer recordings to produce singer-specific variance encodings using contrastive learning. We subsequently trained a well-known autoencoder framework (AutoVC) conditioned on these SIEs, and measured differences in SVC performance when using different latent regressor loss components. We found that using this loss w.r.t. SIEs leads to better performance than w.r.t. bottleneck embeddings, where converted audio is more natural and specific towards target singers. The inclusion of this loss component has the advantage of explicitly forcing the network to reconstruct with timbral similarity, and also negates the effect of poor disentanglement in AutoVC's bottleneck embeddings. We demonstrate peculiar diversity between computational and human evaluations on singer-converted audio clips, which highlights the necessity of both. We also propose a pitch-matching mechanism between source and target singers to ensure these evaluations are not influenced by differences in pitch register.

Yin-Jyun Luo, Kin Wai Cheuk, Woosung Choi et al.

Existing work on pitch and timbre disentanglement has been mostly focused on single-instrument music audio, excluding the cases where multiple instruments are presented. To fill the gap, we propose DisMix, a generative framework in which the pitch and timbre representations act as modular building blocks for constructing the melody and instrument of a source, and the collection of which forms a set of per-instrument latent representations underlying the observed mixture. By manipulating the representations, our model samples mixtures with novel combinations of pitch and timbre of the constituent instruments. We can jointly learn the disentangled pitch-timbre representations and a latent diffusion transformer that reconstructs the mixture conditioned on the set of source-level representations. We evaluate the model using both a simple dataset of isolated chords and a realistic four-part chorales in the style of J.S. Bach, identify the key components for the success of disentanglement, and demonstrate the application of mixture transformation based on source-level attribute manipulation.

Drew Edwards, Xavier Riley, Pedro Sarmento et al.

Guitar tablatures enrich the structure of traditional music notation by assigning each note to a string and fret of a guitar in a particular tuning, indicating precisely where to play the note on the instrument. The problem of generating tablature from a symbolic music representation involves inferring this string and fret assignment per note across an entire composition or performance. On the guitar, multiple string-fret assignments are possible for most pitches, which leads to a large combinatorial space that prevents exhaustive search approaches. Most modern methods use constraint-based dynamic programming to minimize some cost function (e.g.\ hand position movement). In this work, we introduce a novel deep learning solution to symbolic guitar tablature estimation. We train an encoder-decoder Transformer model in a masked language modeling paradigm to assign notes to strings. The model is first pre-trained on DadaGP, a dataset of over 25K tablatures, and then fine-tuned on a curated set of professionally transcribed guitar performances. Given the subjective nature of assessing tablature quality, we conduct a user study amongst guitarists, wherein we ask participants to rate the playability of multiple versions of tablature for the same four-bar excerpt. The results indicate our system significantly outperforms competing algorithms.

Zixun Guo, Simon Dixon

Moonbeam is a transformer-based foundation model for symbolic music, pretrained on a large and diverse collection of MIDI data totaling 81.6K hours of music and 18 billion tokens. Moonbeam incorporates music-domain inductive biases by capturing both absolute and relative musical attributes through the introduction of a novel domain-knowledge-inspired tokenization method and Multidimensional Relative Attention (MRA), which captures relative music information without additional trainable parameters. Leveraging the pretrained Moonbeam, we propose 2 finetuning architectures with full anticipatory capabilities, targeting 2 categories of downstream tasks: symbolic music understanding and conditional music generation (including music infilling). Our model outperforms other large-scale pretrained music models in most cases in terms of accuracy and F1 score across 3 downstream music classification tasks on 4 datasets. Moreover, our finetuned conditional music generation model outperforms a strong transformer baseline with a REMI-like tokenizer. We open-source the code, pretrained model, and generated samples on Github.

Yixiao Zhang, Yukara Ikemiya, Gus Xia et al. · bytedance

Recent advances in text-to-music generation models have opened new avenues in musical creativity. However, music generation usually involves iterative refinements, and how to edit the generated music remains a significant challenge. This paper introduces a novel approach to the editing of music generated by such models, enabling the modification of specific attributes, such as genre, mood and instrument, while maintaining other aspects unchanged. Our method transforms text editing to \textit{latent space manipulation} while adding an extra constraint to enforce consistency. It seamlessly integrates with existing pretrained text-to-music diffusion models without requiring additional training. Experimental results demonstrate superior performance over both zero-shot and certain supervised baselines in style and timbre transfer evaluations. Additionally, we showcase the practical applicability of our approach in real-world music editing scenarios.

Xavier Riley, Drew Edwards, Simon Dixon

Automatic music transcription (AMT) has achieved high accuracy for piano due to the availability of large, high-quality datasets such as MAESTRO and MAPS, but comparable datasets are not yet available for other instruments. In recent work, however, it has been demonstrated that aligning scores to transcription model activations can produce high quality AMT training data for instruments other than piano. Focusing on the guitar, we refine this approach to training on score data using a dataset of commercially available score-audio pairs. We propose the use of a high-resolution piano transcription model to train a new guitar transcription model. The resulting model obtains state-of-the-art transcription results on GuitarSet in a zero-shot context, improving on previously published methods.

Huan Zhang, Taegyun Kwon, Anders Friburg et al.

This paper presents a comprehensive documentation of RenCon 2025, the revival of the expressive performance rendering competition which took place at ISMIR 2025 in Daejeon, Korea. The competition attracted 9 entries from international research groups, representing diverse approaches to expressive piano performance rendering. The two-phase assessment structure comprised a preliminary online evaluation and live real-time rendering at the conference. We analyze the competition format, participant demographics, system performance, and lessons learned for future iterations. The results demonstrate significant advances in expressive rendering capabilities while highlighting remaining challenges in achieving human-level musical expression.

Drew Edwards, Simon Dixon, Emmanouil Benetos et al.

Algorithms for automatic piano transcription have improved dramatically in recent years due to new datasets and modeling techniques. Recent developments have focused primarily on adapting new neural network architectures, such as the Transformer and Perceiver, in order to yield more accurate systems. In this work, we study transcription systems from the perspective of their training data. By measuring their performance on out-of-distribution annotated piano data, we show how these models can severely overfit to acoustic properties of the training data. We create a new set of audio for the MAESTRO dataset, captured automatically in a professional studio recording environment via Yamaha Disklavier playback. Using various data augmentation techniques when training with the original and re-performed versions of the MAESTRO dataset, we achieve state-of-the-art note-onset accuracy of 88.4 F1-score on the MAPS dataset, without seeing any of its training data. We subsequently analyze these data augmentation techniques in a series of ablation studies to better understand their influence on the resulting models.

Sungkyun Chang, Simon Dixon, Emmanouil Benetos

This study introduces RUMAA, a transformer-based framework for music performance analysis that unifies score-to-performance alignment, score-informed transcription, and mistake detection in a near end-to-end manner. Unlike prior methods addressing these tasks separately, RUMAA integrates them using pre-trained score and audio encoders and a novel tri-stream decoder capturing task interdependencies through proxy tasks. It aligns human-readable MusicXML scores with repeat symbols to full-length performance audio, overcoming traditional MIDI-based methods that rely on manually unfolded score-MIDI data with pre-specified repeat structures. RUMAA matches state-of-the-art alignment methods on non-repeated scores and outperforms them on scores with repeats in a public piano music dataset, while also delivering promising transcription and mistake detection results.

Brendan O'Connor, Simon Dixon, George Fazekas

In this paper we propose modifications to the neural network framework, AutoVC for the task of singing technique conversion. This includes utilising a pretrained singing technique encoder which extracts technique information, upon which a decoder is conditioned during training. By swapping out a source singer's technique information for that of the target's during conversion, the input spectrogram is reconstructed with the target's technique. We document the beneficial effects of omitting the latent loss, the importance of sequential training, and our process for fine-tuning the bottleneck. We also conducted a listening study where participants rate the specificity of technique-converted voices as well as their naturalness. From this we are able to conclude how effective the technique conversions are and how different conditions affect them, while assessing the model's ability to reconstruct its input data.

Brendan O'Connor, Simon Dixon, George Fazekas

Sixty participants provided dissimilarity ratings between various singing techniques. Multidimensional scaling, class averaging and clustering techniques were used to analyse timbral spaces and how they change between different singers, genders and registers. Clustering analysis showed that ground-truth similarity and silhouette scores that were not significantly different between gender or register conditions, while similarity scores were positively correlated with participants' instrumental abilities and task comprehension. Participant feedback showed how a revised study design might mitigate noise in our data, leading to more detailed statistical results. Timbre maps and class distance analysis showed us which singing techniques remained similar to one another across gender and register conditions. This research provides insight into how the timbre space of singing changes under different conditions, highlights the subjectivity of perception between participants, and provides generalised timbre maps for regularisation in machine learning.

Emir Demirel, Sven Ahlbäck, Simon Dixon

This paper makes several contributions to automatic lyrics transcription (ALT) research. Our main contribution is a novel variant of the Multistreaming Time-Delay Neural Network (MTDNN) architecture, called MSTRE-Net, which processes the temporal information using multiple streams in parallel with varying resolutions keeping the network more compact, and thus with a faster inference and an improved recognition rate than having identical TDNN streams. In addition, two novel preprocessing steps prior to training the acoustic model are proposed. First, we suggest using recordings from both monophonic and polyphonic domains during training the acoustic model. Second, we tag monophonic and polyphonic recordings with distinct labels for discriminating non-vocal silence and music instances during alignment. Moreover, we present a new test set with a considerably larger size and a higher musical variability compared to the existing datasets used in ALT literature, while maintaining the gender balance of the singers. Our best performing model sets the state-of-the-art in lyrics transcription by a large margin. For reproducibility, we publicly share the identifiers to retrieve the data used in this paper.

Carlos Lordelo, Emmanouil Benetos, Simon Dixon et al.

This paper proposes a deep convolutional neural network for performing note-level instrument assignment. Given a polyphonic multi-instrumental music signal along with its ground truth or predicted notes, the objective is to assign an instrumental source for each note. This problem is addressed as a pitch-informed classification task where each note is analysed individually. We also propose to utilise several kernel shapes in the convolutional layers in order to facilitate learning of efficient timbre-discriminative feature maps. Experiments on the MusicNet dataset using 7 instrument classes show that our approach is able to achieve an average F-score of 0.904 when the original multi-pitch annotations are used as the pitch information for the system, and that it also excels if the note information is provided using third-party multi-pitch estimation algorithms. We also include ablation studies investigating the effects of the use of multiple kernel shapes and comparing different input representations for the audio and the note-related information.

Emir Demirel, Sven Ahlback, Simon Dixon

Recent automatic lyrics transcription (ALT) approaches focus on building stronger acoustic models or in-domain language models, while the pronunciation aspect is seldom touched upon. This paper applies a novel computational analysis on the pronunciation variances in sung utterances and further proposes a new pronunciation model adapted for singing. The singing-adapted model is tested on multiple public datasets via word recognition experiments. It performs better than the standard speech dictionary in all settings reporting the best results on ALT in a capella recordings using n-gram language models. For reproducibility, we share the sentence-level annotations used in testing, providing a new benchmark evaluation set for ALT.

Emir Demirel, Sven Ahlbäck, Simon Dixon

Lyrics alignment in long music recordings can be memory exhaustive when performed in a single pass. In this study, we present a novel method that performs audio-to-lyrics alignment with a low memory consumption footprint regardless of the duration of the music recording. The proposed system first spots the anchoring words within the audio signal. With respect to these anchors, the recording is then segmented and a second-pass alignment is performed to obtain the word timings. We show that our audio-to-lyrics alignment system performs competitively with the state-of-the-art, while requiring much less computational resources. In addition, we utilise our lyrics alignment system to segment the music recordings into sentence-level chunks. Notably on the segmented recordings, we report the lyrics transcription scores on a number of benchmark test sets. Finally, our experiments highlight the importance of the source separation step for good performance on the transcription and alignment tasks. For reproducibility, we publicly share our code with the research community.

Ruchit Agrawal, Daniel Wolff, Simon Dixon

The identification of structural differences between a music performance and the score is a challenging yet integral step of audio-to-score alignment, an important subtask of music information retrieval. We present a novel method to detect such differences between the score and performance for a given piece of music using progressively dilated convolutional neural networks. Our method incorporates varying dilation rates at different layers to capture both short-term and long-term context, and can be employed successfully in the presence of limited annotated data. We conduct experiments on audio recordings of real performances that differ structurally from the score, and our results demonstrate that our models outperform standard methods for structure-aware audio-to-score alignment.

Carlos Lordelo, Emmanouil Benetos, Simon Dixon et al.

This paper addresses the problem of domain adaptation for the task of music source separation. Using datasets from two different domains, we compare the performance of a deep learning-based harmonic-percussive source separation model under different training scenarios, including supervised joint training using data from both domains and pre-training in one domain with fine-tuning in another. We propose an adversarial unsupervised domain adaptation approach suitable for the case where no labelled data (ground-truth source signals) from a target domain is available. By leveraging unlabelled data (only mixtures) from this domain, experiments show that our framework can improve separation performance on the new domain without losing any considerable performance on the original domain. The paper also introduces the Tap & Fiddle dataset, a dataset containing recordings of Scandinavian fiddle tunes along with isolated tracks for 'foot-tapping' and 'violin'.

Ruchit Agrawal, Simon Dixon

Audio-to-score alignment aims at generating an accurate mapping between a performance audio and the score of a given piece. Standard alignment methods are based on Dynamic Time Warping (DTW) and employ handcrafted features, which cannot be adapted to different acoustic conditions. We propose a method to overcome this limitation using learned frame similarity for audio-to-score alignment. We focus on offline audio-to-score alignment of piano music. Experiments on music data from different acoustic conditions demonstrate that our method achieves higher alignment accuracy than a standard DTW-based method that uses handcrafted features, and generates robust alignments whilst being adaptable to different domains at the same time.

Ruchit Agrawal, Simon Dixon

Audio-to-score alignment aims at generating an accurate mapping between a performance audio and the score of a given piece. Standard alignment methods are based on Dynamic Time Warping (DTW) and employ handcrafted features. We explore the usage of neural networks as a preprocessing step for DTW-based automatic alignment methods. Experiments on music data from different acoustic conditions demonstrate that this method generates robust alignments whilst being adaptable at the same time.

Emir Demirel, Sven Ahlback, Simon Dixon

Speech recognition is a well developed research field so that the current state of the art systems are being used in many applications in the software industry, yet as by today, there still does not exist such robust system for the recognition of words and sentences from singing voice. This paper proposes a complete pipeline for this task which may commonly be referred as automatic lyrics transcription (ALT). We have trained convolutional time-delay neural networks with self-attention on monophonic karaoke recordings using a sequence classification objective for building the acoustic model. The dataset used in this study, DAMP - Sing! 300x30x2 [1] is filtered to have songs with only English lyrics. Different language models are tested including MaxEnt and Recurrent Neural Networks based methods which are trained on the lyrics of pop songs in English. An in-depth analysis of the self-attention mechanism is held while tuning its context width and the number of attention heads. Using the best settings, our system achieves notable improvement to the state-of-the-art in ALT and provides a new baseline for the task.

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.

Daniel Stoller, Mi Tian, Sebastian Ewert et al.

Convolutional neural networks (CNNs) with dilated filters such as the Wavenet or the Temporal Convolutional Network (TCN) have shown good results in a variety of sequence modelling tasks. However, efficiently modelling long-term dependencies in these sequences is still challenging. Although the receptive field of these models grows exponentially with the number of layers, computing the convolutions over very long sequences of features in each layer is time and memory-intensive, prohibiting the use of longer receptive fields in practice. To increase efficiency, we make use of the "slow feature" hypothesis stating that many features of interest are slowly varying over time. For this, we use a U-Net architecture that computes features at multiple time-scales and adapt it to our auto-regressive scenario by making convolutions causal. We apply our model ("Seq-U-Net") to a variety of tasks including language and audio generation. In comparison to TCN and Wavenet, our network consistently saves memory and computation time, with speed-ups for training and inference of over 4x in the audio generation experiment in particular, while achieving a comparable performance in all tasks.

Daniel Stoller, Sebastian Ewert, Simon Dixon

Generative adversarial networks (GANs) have shown great success in applications such as image generation and inpainting. However, they typically require large datasets, which are often not available, especially in the context of prediction tasks such as image segmentation that require labels. Therefore, methods such as the CycleGAN use more easily available unlabelled data, but do not offer a way to leverage additional labelled data for improved performance. To address this shortcoming, we show how to factorise the joint data distribution into a set of lower-dimensional distributions along with their dependencies. This allows splitting the discriminator in a GAN into multiple "sub-discriminators" that can be independently trained from incomplete observations. Their outputs can be combined to estimate the density ratio between the joint real and the generator distribution, which enables training generators as in the original GAN framework. We apply our method to image generation, image segmentation and audio source separation, and obtain improved performance over a standard GAN when additional incomplete training examples are available. For the Cityscapes segmentation task in particular, our method also improves accuracy by an absolute 14.9% over CycleGAN while using only 25 additional paired examples.

Carlos Lordelo, Emmanouil Benetos, Simon Dixon et al.

In this paper we propose an efficient deep learning encoder-decoder network for performing Harmonic-Percussive Source Separation (HPSS). It is shown that we are able to greatly reduce the number of model trainable parameters by using a dense arrangement of skip connections between the model layers. We also explore the utilisation of different kernel sizes for the 2D filters of the convolutional layers with the objective of allowing the network to learn the different time-frequency patterns associated with percussive and harmonic sources more efficiently. The training and evaluation of the separation has been done using the training and test sets of the MUSDB18 dataset. Results show that the proposed deep network achieves automatic learning of high-level features and maintains HPSS performance at a state-of-the-art level while reducing the number of parameters and training time.

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.

Daniel Stoller, Sebastian Ewert, Simon Dixon

Models for audio source separation usually operate on the magnitude spectrum, which ignores phase information and makes separation performance dependant on hyper-parameters for the spectral front-end. Therefore, we investigate end-to-end source separation in the time-domain, which allows modelling phase information and avoids fixed spectral transformations. Due to high sampling rates for audio, employing a long temporal input context on the sample level is difficult, but required for high quality separation results because of long-range temporal correlations. In this context, we propose the Wave-U-Net, an adaptation of the U-Net to the one-dimensional time domain, which repeatedly resamples feature maps to compute and combine features at different time scales. We introduce further architectural improvements, including an output layer that enforces source additivity, an upsampling technique and a context-aware prediction framework to reduce output artifacts. Experiments for singing voice separation indicate that our architecture yields a performance comparable to a state-of-the-art spectrogram-based U-Net architecture, given the same data. Finally, we reveal a problem with outliers in the currently used SDR evaluation metrics and suggest reporting rank-based statistics to alleviate this problem.

Daniel Stoller, Sebastian Ewert, Simon Dixon

A main challenge in applying deep learning to music processing is the availability of training data. One potential solution is Multi-task Learning, in which the model also learns to solve related auxiliary tasks on additional datasets to exploit their correlation. While intuitive in principle, it can be challenging to identify related tasks and construct the model to optimally share information between tasks. In this paper, we explore vocal activity detection as an additional task to stabilise and improve the performance of vocal separation. Further, we identify problematic biases specific to each dataset that could limit the generalisation capability of separation and detection models, to which our proposed approach is robust. Experiments show improved performance in separation as well as vocal detection compared to single-task baselines. However, we find that the commonly used Signal-to-Distortion Ratio (SDR) metrics did not capture the improvement on non-vocal sections, indicating the need for improved evaluation methodologies.

Adib Mehrabi, Keunwoo Choi, Simon Dixon et al.

The expressive nature of the voice provides a powerful medium for communicating sonic ideas, motivating recent research on methods for query by vocalisation. Meanwhile, deep learning methods have demonstrated state-of-the-art results for matching vocal imitations to imitated sounds, yet little is known about how well learned features represent the perceptual similarity between vocalisations and queried sounds. In this paper, we address this question using similarity ratings between vocal imitations and imitated drum sounds. We use a linear mixed effect regression model to show how features learned by convolutional auto-encoders (CAEs) perform as predictors for perceptual similarity between sounds. Our experiments show that CAEs outperform three baseline feature sets (spectrogram-based representations, MFCCs, and temporal features) at predicting the subjective similarity ratings. We also investigate how the size and shape of the encoded layer effects the predictive power of the learned features. The results show that preservation of temporal information is more important than spectral resolution for this application.

Daniel Stoller, Sebastian Ewert, Simon Dixon

The state of the art in music source separation employs neural networks trained in a supervised fashion on multi-track databases to estimate the sources from a given mixture. With only few datasets available, often extensive data augmentation is used to combat overfitting. Mixing random tracks, however, can even reduce separation performance as instruments in real music are strongly correlated. The key concept in our approach is that source estimates of an optimal separator should be indistinguishable from real source signals. Based on this idea, we drive the separator towards outputs deemed as realistic by discriminator networks that are trained to tell apart real from separator samples. This way, we can also use unpaired source and mixture recordings without the drawbacks of creating unrealistic music mixtures. Our framework is widely applicable as it does not assume a specific network architecture or number of sources. To our knowledge, this is the first adoption of adversarial training for music source separation. In a prototype experiment for singing voice separation, separation performance increases with our approach compared to purely supervised training.

Eita Nakamura, Kazuyoshi Yoshii, Simon Dixon

This paper presents a statistical method for use in music transcription that can estimate score times of note onsets and offsets from polyphonic MIDI performance signals. Because performed note durations can deviate largely from score-indicated values, previous methods had the problem of not being able to accurately estimate offset score times (or note values) and thus could only output incomplete musical scores. Based on observations that the pitch context and onset score times are influential on the configuration of note values, we construct a context-tree model that provides prior distributions of note values using these features and combine it with a performance model in the framework of Markov random fields. Evaluation results show that our method reduces the average error rate by around 40 percent compared to existing/simple methods. We also confirmed that, in our model, the score model plays a more important role than the performance model, and it automatically captures the voice structure by unsupervised learning.

Siying Wang, Sebastian Ewert, Simon Dixon

Large music content libraries often comprise multiple versions of a piece of music. To establish a link between different versions, automatic music alignment methods map each position in one version to a corresponding position in another version. Due to the leeway in interpreting a piece, any two versions can differ significantly, for example, in terms of local tempo, articulation, or playing style. For a given pair of versions, these differences can be significant such that even state-of-the-art methods fail to identify a correct alignment. In this paper, we present a novel method that increases the robustness for difficult to align cases. Instead of aligning only pairs of versions as done in previous methods, our method aligns multiple versions in a joint manner. This way, the alignment can be computed by comparing each version not only with one but with several versions, which stabilizes the comparison and leads to an increase in alignment robustness. Using recordings from the Mazurka Project, the alignment error for our proposed method was 14% lower on average compared to a state-of-the-art method, with significantly less outliers (standard deviation 53% lower).

Siddharth Sigtia, Emmanouil Benetos, Simon Dixon

We present a supervised neural network model for polyphonic piano music transcription. The architecture of the proposed model is analogous to speech recognition systems and comprises an acoustic model and a music language model. The acoustic model is a neural network used for estimating the probabilities of pitches in a frame of audio. The language model is a recurrent neural network that models the correlations between pitch combinations over time. The proposed model is general and can be used to transcribe polyphonic music without imposing any constraints on the polyphony. The acoustic and language model predictions are combined using a probabilistic graphical model. Inference over the output variables is performed using the beam search algorithm. We perform two sets of experiments. We investigate various neural network architectures for the acoustic models and also investigate the effect of combining acoustic and music language model predictions using the proposed architecture. We compare performance of the neural network based acoustic models with two popular unsupervised acoustic models. Results show that convolutional neural network acoustic models yields the best performance across all evaluation metrics. We also observe improved performance with the application of the music language models. Finally, we present an efficient variant of beam search that improves performance and reduces run-times by an order of magnitude, making the model suitable for real-time applications.

Siddharth Sigtia, Emmanouil Benetos, Nicolas Boulanger-Lewandowski et al.

We investigate the problem of incorporating higher-level symbolic score-like information into Automatic Music Transcription (AMT) systems to improve their performance. We use recurrent neural networks (RNNs) and their variants as music language models (MLMs) and present a generative architecture for combining these models with predictions from a frame level acoustic classifier. We also compare different neural network architectures for acoustic modeling. The proposed model computes a distribution over possible output sequences given the acoustic input signal and we present an algorithm for performing a global search for good candidate transcriptions. The performance of the proposed model is evaluated on piano music from the MAPS dataset and we observe that the proposed model consistently outperforms existing transcription methods.

Peter Foster, Simon Dixon, Anssi Klapuri

This paper investigates methods for quantifying similarity between audio signals, specifically for the task of of cover song detection. We consider an information-theoretic approach, where we compute pairwise measures of predictability between time series. We compare discrete-valued approaches operating on quantised audio features, to continuous-valued approaches. In the discrete case, we propose a method for computing the normalised compression distance, where we account for correlation between time series. In the continuous case, we propose to compute information-based measures of similarity as statistics of the prediction error between time series. We evaluate our methods on two cover song identification tasks using a data set comprised of 300 Jazz standards and using the Million Song Dataset. For both datasets, we observe that continuous-valued approaches outperform discrete-valued approaches. We consider approaches to estimating the normalised compression distance (NCD) based on string compression and prediction, where we observe that our proposed normalised compression distance with alignment (NCDA) improves average performance over NCD, for sequential compression algorithms. Finally, we demonstrate that continuous-valued distances may be combined to improve performance with respect to baseline approaches. Using a large-scale filter-and-refine approach, we demonstrate state-of-the-art performance for cover song identification using the Million Song Dataset.

Peter Foster, Matthias Mauch, Simon Dixon

We propose string compressibility as a descriptor of temporal structure in audio, for the purpose of determining musical similarity. Our descriptors are based on computing track-wise compression rates of quantised audio features, using multiple temporal resolutions and quantisation granularities. To verify that our descriptors capture musically relevant information, we incorporate our descriptors into similarity rating prediction and song year prediction tasks. We base our evaluation on a dataset of 15500 track excerpts of Western popular music, for which we obtain 7800 web-sourced pairwise similarity ratings. To assess the agreement among similarity ratings, we perform an evaluation under controlled conditions, obtaining a rank correlation of 0.33 between intersected sets of ratings. Combined with bag-of-features descriptors, we obtain performance gains of 31.1% and 10.9% for similarity rating prediction and song year prediction. For both tasks, analysis of selected descriptors reveals that representing features at multiple time scales benefits prediction accuracy.