Music was the first mass-market industry to be completely restructured by digital technology, and today we can have access to thousands of tracks stored locally on our smartphone and millions of tracks through cloud-based music services. Given the vast quantity of music at our fingertips, we now require novel ways of describing, indexing, searching and interacting with musical content. In this thesis we focus on a technology that opens the door to a wide range of such applications: automatically estimating the pitch sequence of the melody directly from the audio signal of a polyphonic music recording, also referred to as melody extraction. Whilst identifying the pitch of the melody is something human listeners can do quite well, doing this automatically is highly challenging. We present a novel method for melody extraction based on the tracking and characterisation of the pitch ...

Much like natural language, music is highly structured, with strong priors on the likelihood of note sequences. In automatic speech recognition (ASR), these priors are called language models, which are used in addition to acoustic models and participate greatly to the success of today's systems. However, in Automatic Music Transcription (AMT), ASR's musical equivalent, Music Language Models (MLMs) are rarely used. AMT can be defined as the process of extracting a symbolic representation from an audio signal, describing which notes were played at what time. In this thesis, we investigate the design of MLMs using recurrent neural networks (RNNs) and their use for AMT. We first look into MLM performance on a polyphonic prediction task. We observe that using musically-relevant timesteps results in desirable MLM behaviour, which is not reflected in usual evaluation metrics. We compare our model against benchmark ...

The burgeoning of deep learning-based music generation has overlooked the potential of symbolic representations tailored for fretted instruments. Guitar tablatures offer an advantageous approach to represent prescriptive information about music performance, often missing from standard MIDI representations. This dissertation tackles a gap in symbolic music generation by developing models that predict both musical structures and expressive guitar performance techniques. We first present DadaGP, a dataset comprising over 25k songs converted from the Guitar Pro tablature format to a dedicated token format suiting sequence models such as the Transformer. To establish a benchmark, we first introduce a baseline unconditional model for guitar tablature generation, by training a Transformer-XL architecture on the DadaGP dataset. We explored various architecture configurations and experimented with two different tokenisation approaches. Delving into controllability of the generative process, we introduce methods for manipulating the output's instrumentation (inst-CTRL) ...

Automatically making sense of digital information, and specially of music digital documents, is an important problem our modern society is facing. In fact, there are still many tasks that, although being easily performed by humans, cannot be effectively performed by a computer. In this work we focus on one of such tasks: the identification of musical piece versions (alternate renditions of the same musical composition like cover songs, live recordings, remixes, etc.). In particular, we adopt a computational approach solely based on the information provided by the audio signal. We propose a system for version identification that is robust to the main musical changes between versions, including timbre, tempo, key and structure changes. Such a system exploits nonlinear time series analysis tools and standard methods for quantitative music description, and it does not make use of a specific modeling strategy ...

Witnessing the omnipresence of ever complex yet so intuitive digital video media, research community has raised the question of its meaningful use and management. Stored in immense multimedia databases, digital videos need to be retrieved and structured in an intelligent way, relying on the content and the rich semantics involved. Therefore, the third generation of Content Based Video Indexing and Retrieval systems faces the problem of the semantic gap between the simplicity of the available visual features and the richness of user semantics. This work focuses on the issues of efficiency and scalability in video indexing and retrieval to facilitate a video representation model capable of semantic annotation. A highly efficient algorithm for temporal analysis and key-frame extraction is developed. It is based on the prediction information extracted directly from the compressed-domain features and the robust scalable analysis in the ...

In this thesis, we examine the problem of algorithmic analysis of complex audio scenes with a special emphasis on natural audio scenes. One of the driving goals behind this work is to develop tools for monitoring the presence of animals in areas of interest based on their vocalisations. This task, which often occurs in the evaluation of nature conservation measures, leads to a number of subproblems in audio scene analysis. In order to develop and evaluate pattern recognition algorithms for animal sounds, a representative collection of such sounds is necessary. Building such a collection is beyond the scope of a single researcher and we therefore use data from the Animal Sound Archive of the Humboldt University of Berlin. Although a large portion of well annotated recordings from this archive has been available in digital form, little infrastructure for searching and ...

In the audio research field, the majority of machine learning systems focus on recognizing a limited number of sound events. However, when a machine interacts with real data, it must be able to handle much more varied and complex situations. To tackle this problem, annotators use natural language, which allows any sound information to be summarized. Automated Audio Captioning (AAC) was introduced recently to develop systems capable of automatically producing a description of any type of sound in text form. This task concerns all kinds of sound events such as environmental, urban, domestic sounds, sound effects, music or speech. This type of system could be used by people who are deaf or hard of hearing, and could improve the indexing of large audio databases. In the first part of this thesis, we present the state of the art of the ...

Music is built from sound, ultimately resulting from an elaborate interaction between the sound-generating properties of physical objects (i.e. music instruments) and the sound perception abilities of the human auditory system. Humans, even without any kind of formal music training, are typically able to ex- tract, almost unconsciously, a great amount of relevant information from a musical signal. Features such as the beat of a musical piece, the main melody of a complex musical ar- rangement, the sound sources and events occurring in a complex musical mixture, the song structure (e.g. verse, chorus, bridge) and the musical genre of a piece, are just some examples of the level of knowledge that a naive listener is commonly able to extract just from listening to a musical piece. In order to do so, the human auditory system uses a variety of cues ...

For humans, the sound is valuable mostly for its meaning. The voice is spoken language, music, artistic intent. Its physiological functioning is highly developed, as well as our understanding of the underlying process. It is a challenge to replicate this analysis using a computer: in many aspects, its capabilities do not match those of human beings when it comes to speech or instruments music recognition from the sound, to name a few. In this thesis, two problems are investigated: the source separation and the musical processing. The first part investigates the source separation using only one Microphone. The problem of sources separation arises when several audio sources are present at the same moment, mixed together and acquired by some sensors (one in our case). In this kind of situation it is natural for a human to separate and to recognize ...

Visual and audio contents always played a key role in communications, because of their immediacy and presumed objectivity. This has become even more true in the digital era, and today it is common to have multimedia contents stand as proof of events. Digital contents, however, are also very easy to manipulate, thus calling for analysis methods devoted to uncover their processing history. Multimedia forensics is the science trying to answer questions about the past of a given image, audio or video file, questions like “which was the recording device?", or “is the content authentic?". In particular, authenticity assessment is a crucial task in many contexts, and it usually consists in determining whether the investigated object has been artificially created by splicing together different contents. In this thesis we address the problem of splicing detection in the three main media: image, ...

Sound source separation refers to the task of estimating the signals produced by individual sound sources from a complex acoustic mixture. It has several applications, since monophonic signals can be processed more efficiently and flexibly than polyphonic mixtures. This thesis deals with the separation of monaural, or, one-channel music recordings. We concentrate on separation methods, where the sources to be separated are not known beforehand. Instead, the separation is enabled by utilizing the common properties of real-world sound sources, which are their continuity, sparseness, and repetition in time and frequency, and their harmonic spectral structures. One of the separation approaches taken here use unsupervised learning and the other uses model-based inference based on sinusoidal modeling. Most of the existing unsupervised separation algorithms are based on a linear instantaneous signal model, where each frame of the input mixture signal is modeled ...

Single channel source separation is a quite recent problem of constantly growing interest in the scientific world. However, this problem is still very far to be solved, and even more, it cannot be solved in all its generality. Indeed, since this problem is highly underdetermined, the main difficulty is that a very strong knowledge about the sources is required to be able to separate them. For a grand class of existing separation methods, this knowledge is expressed by statistical source models, notably Gaussian Mixture Models (GMM), which are learned from some training examples. The subject of this work is to study the separation methods based on statistical models in general, and then to apply them to the particular problem of separating singing voice from background music in mono recordings of songs. It can be very useful to propose some satisfactory ...

In this thesis we address the subject of automatic extraction of harmony information from audio recordings. We focus on chord symbol recognition and methods for evaluating algorithms designed to perform that task. We present a novel six-dimensional model for equal tempered pitch space based on concepts from neo-Riemannian music theory. This model is employed as the basis of a harmonic change detection function which we use to improve the performance of a chord recognition algorithm. We develop a machine readable text syntax for chord symbols and present a hand labelled chord transcription collection of 180 Beatles songs annotated using this syntax. This collection has been made publicly available and is already widely used for evaluation purposes in the research community. We also introduce methods for comparing chord symbols which we subsequently use for analysing the statistics of the transcription collection. ...

Everyday sounds in real-world environments are a powerful source of information by which humans can interact with their environments. Humans can infer what is happening around them by listening to everyday sounds. At the same time, it is a challenging task for a computer algorithm in a smart device to automatically recognise, understand, and interpret everyday sounds. Sound event detection (SED) is the process of transcribing an audio recording into sound event tags with onset and offset time values. This involves classification and segmentation of sound events in the given audio recording. SED has numerous applications in everyday life which include security and surveillance, automation, healthcare monitoring, multimedia information retrieval, and assisted living technologies. SED is to everyday sounds what automatic speech recognition (ASR) is to speech and automatic music transcription (AMT) is to music. The fundamental questions in designing ...

When listening to music, some humans can easily recognize which instruments play at what time or when a new musical segment starts, but cannot describe exactly how they do this. To automatically describe particular aspects of a music piece – be it for an academic interest in emulating human perception, or for practical applications –, we can thus not directly replicate the steps taken by a human. We can, however, exploit that humans can easily annotate examples, and optimize a generic function to reproduce these annotations. In this thesis, I explore solving different music perception tasks with deep learning, a recent branch of machine learning that optimizes functions of many stacked nonlinear operations – referred to as deep neural networks – and promises to obtain better results or require less domain knowledge than more traditional techniques. In particular, I employ ...