Call for PhD applications at the Machine Listening Lab

The Machine Listening Lab is welcoming PhD applications for September 2022 entry. Applicants from all nationalities can apply across different funding schemes and PhD programmes. Current PhD funding opportunities for September 2022 entry include:

Applicants are encouraged to contact prospective supervisors before submitting their application – please send an email to your selected supervisors with your CV and draft research proposal.

Suggested PhD topics offered by Machine Listening Lab academics include:


Computational analysis of chick vocalisations: from categorisation to live feedback
(for PhD in Computer Science funded by a QMUL Principal’s studentship)

Supervisors: Dr. Emmanouil Benetos and Dr. Elisabetta Versace

The assessment of animals’ emotional state and welfare is a central issue for behavioural neuroscience and ethical farming. Animal vocalisations provide a rich set of information on the inner state of animals and can be used to influence animal behaviour in industrial settings, such as chicken farms. However, there is limited research on using vocalisations for monitoring animal welfare, and on the use of audio technologies for automatic welfare assessment of poultry in industrial settings.

In this project, we will develop computational methods to automatically categorise vocalisations of domestic chickens, infer their emotional state, provide live feedback and identify stimuli that can improve animal welfare. The PhD project will build upon pilot work led by the supervisors [i] on automatic recognition of chick calls using machine learning and signal processing methods. This project will lead to computationally efficient and robust machine learning methods and systems for automatically monitoring poultry welfare from audio, as well as will investigate research questions related to poultry development, behaviour, and well-being in industrial settings. Prospective candidates should be curious, self-motivated and have experience in one or more of the following: Bioacoustics, Cognitive Science, Artificial Intelligence/Machine Learning, Digital Signal Processing.

[i] C. Wang, E. Benetos, S. Wang, and E. Versace, “Joint Scattering for Automatic Chick Call Recognition”, 2022 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), submitted.


Discourse structure recognition for broadcast summarisation
(for PhD in Computer Science funded by a China Scholarship Council studentship)

Supervisors: Prof. Matthew Purver, Dr Huy Phan

This project will investigate new methods for inferring and detecting dialogue and narrative structure in broadcast data, primarily TV programmes, to enable more effective and meaningful semantic search and linking, and/or generation of concise meaningful summaries. Current methods for search and summarization are based around models of word and phrase meaning: these are effective when the key information is expressed though verbal content, and our recent work has shown that e.g. news broadcast segmentation can be improved by incorporating recent advances in NLP methods (Ghinassi, DataTV 2021). However, in many domains this is not the case: political debates need to be understood as networks of linked but contrasting opinions; dramas as often non-linear plotlines with key events and changes. Effective models of these formats must therefore include information about these elements of narrative/discourse structure, and learn to detect these in real datasets.

Methods for recognising and inferring narrative structure are available from related work on text understanding and summarization (including work from our lab, Droog-Hayes et al, ICSC 2019). This project will extend them to be suitable for spoken broadcast data and to use state-of-the-art neural network NLP methods, following our recent work in topic segmentation (Ghinassi, 2021) and related recent work in text segmentation (Angelov, 2020; Lukasik et al., 2021), while integrating information from multiple modalities including text, audio and video, again following our recent advances in multimodal modelling (Ghinassi et al., in prep.; Rohanian et al., Interspeech 2020). It will then build on this to develop models for jointly learning to generate compressed summaries while detecting key events and points in the discourse (discussion points, key characters, plot changes etc.), by adapting recent neural methods for dialogue act detection and summarization (Goo & Chen, 2018; Goo et al., 2018), and combining with suitable graph structures


Multitask modelling for overlapping sound sources
(for PhD in Artificial Intelligence and Music)

Supervisor: Dr Huy Phan

Overlapping sound sources are the main error source in a modelling system. Examples are polyphonic audio events in audio event detection and polyphonic music in multi-instrument transcription. In deep-learning context, the most common approach to deal with event overlaps is to treat the modelling task as a multi-label classification problem. By doing this, we inherently consider multiple one-vs.-rest classification problems, which are jointly solved by a single (i.e. shared) network. This project investigates to frame the task as a multi-class classification problem by considering each possible label combination as one class. To circumvent the large number of arising classes due to combinatorial explosion, decomposition of the label space will be explored to form multiple groups of category labels and yield a multi-task problem in a divide-and-conquer fashion, where each of the tasks is a multi-class classification problem. Network architectures will be then devised for multi-task modelling. The approach will be validated on databases with high overlapping degree of sound sources for polyphonic audio event detection and polyphonic music transcription.


Personalized scientific-grading sleep monitoring in home environments
(for PhD in Computer Science funded by a China Scholarship Council studentship)

Supervisor: Dr Huy Phan

Good sleep is crucial in maintaining one’s mental and physical health while sleep disorders are linked with plethora of different ailments, such as cardiovascular diseases, dementia, and depression to name a few. Accurate and cost-effective monitoring of sleep not only has great medical value but also allows individuals to self-assess and self-manage their sleep. Its importance gives rise to increasing demand in bringing sleep monitoring from labs to home environments for longitudinal monitoring. Existing commercial devices has limited value for scientific purposes, such as sleep disorders assessment or studying sleep under conditions like dementia. Novel wearable in-ear EEG devices, whose sensors are fitted neatly inside an ear canal to measure brain activities, hold a great potential for home-based use. Furthermore, they have been shown to be comparable to the lab-based polysomnography for sleep scoring.

However, there are two data-related challenges with these devices. First, manually labelling a large amount of their data is difficult and expensive as it requires to have polysomnography data recorded in parallel. Second, it exhibits strong “trait-like characteristics” specific to individuals and significant night-to-night variation in an individual’s sleep. This project aims to develop personalized deep learning methods to overcome these challenges. Domain adaptation methods will be explored to transfer knowledge from a large polysomnography database to ear-EEG. Regularization techniques will be developed given a potentially small amount of labelled ear-EEG data. Semi-supervised/unsupervised domain adaptation will be also explored to leverage unlabelled data. Furthermore, we will also investigate continual learning methods to keep up a personalized model with the possible “concept drift” caused by the long-term changes in a person’s sleep patterns. While these methods are required to learn from sequential, and potentially small, data, they also need to overcome catastrophic forgetting.


Resource-efficient models for music understanding
(for PhD in Artificial Intelligence and Music)

Supervisors: Dr. Emmanouil Benetos and Prof. Phillip Stanley-Marbell

State-of-the-art models for music understanding and music information research are often very hard to run on small and embedded devices such as mobile phones, single-board computers, and other microprocessors. At the same time, the computational cost, footprint, and environmental impact for building and deploying deep learning models for music understanding is constantly increasing. This PhD project will investigate methods for creating resource-efficient models for music understanding, applied to various tasks in music information research that involve music audio data, such as automatic music transcription, audio fingerprinting, or music tagging. Methods to be investigated can include but are not limited to sparse training, network pruning, binary neural networks, post-training inference, and knowledge distillation.

The successful candidate will investigate, propose and develop novel machine learning methods and software tools for resource-efficient music understanding, and will apply them to address tasks of their choice within the wider field of music information research. This will result in models that can be deployed on small or embedded devices, or on offline models where learning and inference times and computational resources are drastically reduced.


Self-supervision in machine listening
(for PhD in Artificial Intelligence and Music)

Supervisor: Dr. Emmanouil Benetos
in collaboration with Bytedance

Self-supervised learning methods aim to provide an alternative to supervised representation learning, eliminating the need for large annotated datasets. Self-supervision has advanced rapidly in recent years with applications across several modalities, and can be ideally used in machine listening and music understanding tasks which have been historically data-deprived compared to other domains. This PhD project will investigate methods for self-supervised learning applied to various tasks in applied to various tasks in music information research that involve music audio data, such as automatic music transcription, audio fingerprinting, or music tagging. Methods to be investigated can include but are not limited to contrastive self-supervised learning, formulation of appropriate pretext tasks, transferability to downstream tasks, and links between self-supervised and semi-supervised learning for music understanding.

The successful candidate will investigate, propose and develop novel self-supervised representation learning methods and software tools for music understanding, and will apply them to address tasks of their choice within the wider field of music information research. This will result in models that can learn from unlabelled data while performing comparably or surpassing supervised learning methods.


Sound source separation and localisation
(for PhD in Computer Science funded by a China Scholarship Council studentship)

Supervisors: Dr. Emmanouil Benetos and Prof. Mark Sandler

Audio and music source separation has been an active area of research, with applications in sound recording & production, broadcasting, and audio consumption. In recent years, deep learning techniques have dominated the topic, along with the use of priors for informed source separation. While the research community has mostly focused on single-channel source separation, real-world, professional studio applications make use of several microphones which are used in consort and their outputs are combined (or mixed) to create a composite signal.

This PhD project will investigate and propose machine learning methods for sound source separation and localisation (SSSL), which will separate a mixture recording into its constituent sources, in the context of a studio setup involving multiple microphones including spot and main feeds, while at the same time using information on the spatial location of sound sources to improve separation performance. The proposed research will lead to a new paradigm for informed sound source separation, where prior information will be provided both by spot mics related to constituent sources, as well as by providing or automatically inferring the spatial location of sound sources in the scene. The proposed project will draw knowledge from and will contribute knowledge to the fields of machine learning, digital signal processing, and acoustics, and will advance the broader fields of audio & music technology and signal separation. Upon completion, the project will provide new high fidelity and computationally efficient algorithms and models for separating and enhancing sources in studio practices.


Using Signal-informed Source Separation (SISS) principals to improve instrument separation from legacy recordings
(for PhD in Artificial Intelligence and Music)

Supervisors: Prof. Mark Sandler and Dr. Emmanouil Benetos

The recently proposed Signal-Informed Source Separation (SISS) paradigm from c4dm belongs to the broader category of Informed Source Separation (ISS), with the unique and specific attribute of using one audio signal to inform the separation of another. The informing source is a close approximation of a coherent component in the mixture. A current AIM PhD is examining this paradigm for live ensemble recordings when the spot mic signal informs the separation of the main mix. An alternative viewpoint is that the informing signal as a caricature of its corresponding component in the main feed. This suggests that we should investigate other caricatures for musical instrument separation and modification, especially for re-mixing and up-mixing of legacy commercial recordings. For example, a session musician could play the guitar line in the Beatles’ “She Loves You” to separate Harrison’s part, or it could be rendered from a MIDI transcription. Preliminary, confirmatory evidence for this approach appears in [1 & 2] which explore crude implementations with good outcomes but do not develop the approach further.

This PhD will develop skills in deep learning, especially architectures employing conditioning, as well as novel cost functions, perhaps incorporating physical models of the instruments to be separated. Applicants would benefit from a background in Machine Learning and DSP, coupled with knowledge of modern music recording, processing and mixing techniques.

[1] P. Smaragdis & G. J. Mysore, ‘Separation by “humming”: User-guided sound extraction from monophonic mixtures’, in IEEE WASPAA, 2009.
[2] Y. Li et al, ‘Learning to Denoise Historical Music’, in ISMIR, 2020.

Posted in Uncategorized