Dissertations / Theses on the topic 'Musical Instrument Recognition'

Trabalho apresentado no âmbito do Mestrado em Engenharia Informática, como requisito parcial para obtenção do grau de Mestre em Engenharia Informática
The subject of automatic indexing of multimedia has been a target of numerous discussion and study. This interest is due to the exponential growth of multimedia content and the subsequent need to create methods that automatically catalogue this data. To fulfil this idea, several projects and areas of study have emerged. The most relevant of these are the MPEG-7 standard, which defines a standardized system for the representation and automatic extraction of information present in the content, and Music Information Retrieval (MIR), which gathers several paradigms and areas of study relating to music. The main approach to this indexing problem relies on analysing data to obtain and identify descriptors that can help define what we intend to recognize (as, for instance,musical instruments, voice, facial expressions, and so on), this then provides us with information we can use to index the data. This dissertation will focus on audio indexing in music, specifically regarding the recognition of musical instruments from recorded musical notes. Moreover, the developed system and techniques will also be tested for the recognition of ambient sounds (such as the sound of running water, cars driving by, and so on). Our approach will use non-negative matrix factorization to extract features from various types of sounds, these will then be used to train a classification algorithm that will be then capable of identifying new sounds.

Thanks to the advancement of technological progress in networking and signal processing, we can access a large amount of musical content. In order for users to search among these vast catalogs, they need to have access to music-related information beyond the pure digital music file. Manual annotation of music is too expensive, therefore automated annotation would be of great use. A meaningful description of the musical pieces requires the incorporation of information about the instruments present in them. In this work, we present an approach for musical instrument recognition using the scattering transform, which is a transformation that gives a translation invariant representation, that is stable to deformations and preserves high frequency information for classication. We study recognition in both singleinstrument and multiple-instrument contexts. We compare the performance of models using the scattering transform to those using other standard features. We also examine the impact of the amount of training data. The experiments carried out do not show a clear superior performance of either feature representation. Still, the scattering transform is worth taking into account when choosing a way to extract features if we want to be able to characterize non-stationary signal structures.
Tack vare den tekniska utvecklingen i nätverk och signalbehandling kan vi få tillgång till en stor mängd musikaliskt innehåll. For att användare ska söka bland dessa stora kataloger måste de ha tillgång till musikrelaterad information utöver den rena digitala musikfilen. Eftersom den manuella annotationsprocessen skulle vara för dyr måste den automatiseras. En meningsfull beskrivning av musikstyckena kräver införlivande av information om instrumenten som finns i dem. I det här arbetet presenterar vi en metod for igenkänning av musikinstrument med hjälp av den scattering transform, som är en transformation som ger en översattnings-invariant representation, som är stabil för deformationer och bevarar högfrekvensinformation för klassicering. Vi studerar igenkännande i både enskilda instrument- och flera instrumentförhållanden. Vi jämför modellerna med den scattering transforms prestanda med de som använder andra standardfunktioner. Vi undersöker också effekterna av mangden traningsdata. Experimenten som utförs visar inte en tydlig överlagsen prestanda for någon av representationsföreställningarna jämfört med den andra. Fortfarande är den scattering transform värd att ta hänsyn till när man väljer ett sätt att extrahera funktioner om vi vill kunna karakterisera icke-stationära signalstrukturer.

En aquesta tesi presentem un mètode general per al reconeixement automàtic d’instruments musicals partint d’un senyal d’àudio. A diferència de molts enfocs relacionats, el nostre evita restriccions artificials o artificioses pel que fa al disseny algorísmic, les dades proporcionades al sistema, o el context d’aplicació. Per tal de fer el problema abordable, limitem el procés a l’operació més bàsica consistent a reconèixer l’instrument predominant en un breu fragment d’àudio. Així ens estalviem la separació de fonts sonores en la mescla i, més específicament, predim una font sonora a partir del timbre general del so analitzat. Per tal de compensar aquesta restricció incorporem, addicionalment, informació derivada d’una anàlisi musical jeràrquica: primer incorporem context temporal a l’hora d’extraure etiquetes dels instruments, després incorporem aspectes formals de la peça que poden ajudar al reconeixement de l’instrument, i finalment incloem informació general gràcies a l’explotació de les associacions entre gèneres musicals i instruments.
In this dissertation we present a method for the automatic recognition of musical instruments from music audio signal. Unlike most related approaches, our specific conception mostly avoids laboratory constraints on the method’s algorithmic design, its input data, or the targeted application context. To account for the complex nature of the input signal, we limit the basic process in the processing chain to the recognition of a single predominant musical instrument from a short audio fragment. We thereby prevent resolving the mixture and rather predict one source from the timbre of the sound. To compensate for this restriction we further incorporate information derived from a hierarchical music analysis; we first incorporate musical context to extract instrumental labels from the time-varying model decisions. Second, the method incorporates information regarding the piece’s formal aspects into the process. Finally, we include information from the collection level by exploiting associations between musical genres and instrumentations.

Thesis (PhD)--Stellenbosch University, 2013.
ENGLISH ABSTRACT: An area of data mining and statistics that is currently receiving considerable attention is the field of multi-label learning. Problems in this field are concerned with scenarios where each data case can be associated with a set of labels instead of only one. In this thesis, we review the field of multi-label learning and discuss the lack of suitable benchmark data available for evaluating multi-label algorithms. We propose a technique for simulating multi-label data, which allows good control over different data characteristics and which could be useful for conducting comparative studies in the multi-label field. We also discuss the explosion in data in recent years, and highlight the need for some form of dimension reduction in order to alleviate some of the challenges presented by working with large datasets. Feature (or variable) selection is one way of achieving dimension reduction, and after a brief discussion of different feature selection techniques, we propose a new technique for feature selection in a multi-label context, based on the concept of independent probes. This technique is empirically evaluated by using simulated multi-label data and it is shown to achieve classification accuracy with a reduced set of features similar to that achieved with a full set of features. The proposed technique for feature selection is then also applied to the field of music information retrieval (MIR), specifically the problem of musical instrument recognition. An overview of the field of MIR is given, with particular emphasis on the instrument recognition problem. The particular goal of (polyphonic) musical instrument recognition is to automatically identify the instruments playing simultaneously in an audio clip, which is not a simple task. We specifically consider the case of duets – in other words, where two instruments are playing simultaneously – and approach the problem as a multi-label classification one. In our empirical study, we illustrate the complexity of musical instrument data and again show that our proposed feature selection technique is effective in identifying relevant features and thereby reducing the complexity of the dataset without negatively impacting on performance.
AFRIKAANSE OPSOMMING: ‘n Area van dataontginning en statistiek wat tans baie aandag ontvang, is die veld van multi-etiket leerteorie. Probleme in hierdie veld beskou scenarios waar elke datageval met ‘n stel etikette geassosieer kan word, instede van slegs een. In hierdie skripsie gee ons ‘n oorsig oor die veld van multi-etiket leerteorie en bespreek die gebrek aan geskikte standaard datastelle beskikbaar vir die evaluering van multi-etiket algoritmes. Ons stel ‘n tegniek vir die simulasie van multi-etiket data voor, wat goeie kontrole oor verskillende data eienskappe bied en wat nuttig kan wees om vergelykende studies in die multi-etiket veld uit te voer. Ons bespreek ook die onlangse ontploffing in data, en beklemtoon die behoefte aan ‘n vorm van dimensie reduksie om sommige van die uitdagings wat deur sulke groot datastelle gestel word die hoof te bied. Veranderlike seleksie is een manier van dimensie reduksie, en na ‘n vlugtige bespreking van verskillende veranderlike seleksie tegnieke, stel ons ‘n nuwe tegniek vir veranderlike seleksie in ‘n multi-etiket konteks voor, gebaseer op die konsep van onafhanklike soek-veranderlikes. Hierdie tegniek word empiries ge-evalueer deur die gebruik van gesimuleerde multi-etiket data en daar word gewys dat dieselfde klassifikasie akkuraatheid behaal kan word met ‘n verminderde stel veranderlikes as met die volle stel veranderlikes. Die voorgestelde tegniek vir veranderlike seleksie word ook toegepas in die veld van musiek dataontginning, spesifiek die probleem van die herkenning van musiekinstrumente. ‘n Oorsig van die musiek dataontginning veld word gegee, met spesifieke klem op die herkenning van musiekinstrumente. Die spesifieke doel van (polifoniese) musiekinstrument-herkenning is om instrumente te identifiseer wat saam in ‘n oudiosnit speel. Ons oorweeg spesifiek die geval van duette – met ander woorde, waar twee instrumente saam speel – en hanteer die probleem as ‘n multi-etiket klassifikasie een. In ons empiriese studie illustreer ons die kompleksiteit van musiekinstrumentdata en wys weereens dat ons voorgestelde veranderlike seleksie tegniek effektief daarin slaag om relevante veranderlikes te identifiseer en sodoende die kompleksiteit van die datastel te verminder sonder ‘n negatiewe impak op klassifikasie akkuraatheid.

Nell'elaborato viene valutata l’efficacia delle features a medio termine spettrali e temporali nel problema del riconoscimento degli strumenti musicali a partire dai loro toni. In particolare ne viene testata la differenza al variare della fonte di dati e la robustezza dell’analisi basata su queste in contesti in cui i campioni provengono da molteplici fonti, con differenti tecniche esecutive e di registrazione. Collateralmente vengono valutate le influenze di altre variabili sull’efficacia del riconoscimento e vengono testate alcune tecniche di analisi complementari.

With the ever growing media and music catalogs, tools that search and navigate this data are important. For more complex search queries, meta-data is needed, but to manually label the vast amounts of new content is impossible. In this thesis, automatic labeling of musical instrument activities in song mixes is investigated, with a focus on ways to alleviate the lack of annotated data for instrument activity detection models. Two methods for alleviating the problem of small amounts of data are proposed and evaluated. Firstly, a self-supervised approach based on automatic labeling and mixing of randomized instrument stems is investigated. Secondly, a domain-adaptation approach that trains models on sampled MIDI files for instrument activity detection on recorded music is explored. The self-supervised approach yields better results compared to the baseline and points to the fact that deep learning models can learn instrument activity detection without an intrinsic musical structure in the audio mix. The domain-adaptation models trained solely on sampled MIDI files performed worse than the baseline, however using MIDI data in conjunction with recorded music boosted the performance. A hybrid model combining both self-supervised learning and domain adaptation by using both sampled MIDI data and recorded music produced the best results overall.
I och med de ständigt växande media- och musikkatalogerna krävs verktyg för att söka och navigera i dessa. För mer komplexa sökförfrågningar så behövs det metadata, men att manuellt annotera de enorma mängderna av ny data är omöjligt. I denna uppsats undersöks automatisk annotering utav instrumentsaktivitet inom musik, med ett fokus på bristen av annoterad data för modellerna för instrumentaktivitetsigenkänning. Två metoder för att komma runt bristen på data föreslås och undersöks. Den första metoden bygger på självövervakad inlärning baserad på automatisk annotering och slumpartad mixning av olika instrumentspår. Den andra metoden använder domänadaption genom att träna modeller på samplade MIDI-filer för detektering av instrument i inspelad musik. Metoden med självövervakning gav bättre resultat än baseline och pekar på att djupinlärningsmodeller kan lära sig instrumentigenkänning trots att ljudmixarna saknar musikalisk struktur. Domänadaptionsmodellerna som endast var tränade på samplad MIDI-data presterade sämre än baseline, men att använda MIDI-data tillsammans med data från inspelad musik gav förbättrade resultat. En hybridmodell som kombinerade både självövervakad inlärning och domänadaption genom att använda både samplad MIDI-data och inspelad musik gav de bästa resultaten totalt.

The purpose of my doctorate work has consisted in the exploration of the potentialities and of the effectiveness of different neural classifiers, by experimenting their application in the solution of classification problems occurring in the fields of interest typical of the research group of the “Laboratorio Circuiti” at the Department of Electronic Engineering in Tor Vergata. Moreover, though inspired by works already developed by other scholars, the adopted neural classifiers have been partially modified, in order to add to them interesting peculiarities not present in the original versions, as well as to adapt them to the applications of interest. These applications can be grouped in two great families. As regards the first application, the objects to be classified are identified by features of static nature, while as regards the second family, the objects to be classified are identified by features evolving in time. In relation to the research fields taken as reference, the ones that belong to the first family are the following: • classification, by means of fuzzy algorithms, of acoustic signals, with the aim of attributing them to the source that generated them (recognition of musical instruments) • exclusive classification of simple human motor acts for the purpose of a precocious diagnosis of nervous system diseases The second family of application has been represented by that research field that aims to the development of neural tools for the Automatic Tanscription of piano pieces. The first part of this thesis has been devoted to the detailed description of the adopted neural classification techniques, as well as of the modifications introduced in order to improve their behavior in relation to the particular applications. In the second part, the experiments by means of which I have estimated the before-mentioned neural classification techniques have been introduced. It exactly deals with experiments carried out in the chosen research fields. For every application, the II results achieved have been reported; in some cases, the further steps to perform have also been proposed. After a brief introduction to the biological neural model, a description follows about the model of the artificial neuron that has afterwards inspired all the other models: the one proposed by McCulloch and Pitts in 1943. Subsequently, the different typologies of architectures that characterize neural networks are shortly introduced, as regards the feed-forward networks as well as the recursive networks. Then, a description of some learning strategies (supervised and unsupervised), adopted in order to train neural networks, is also given; some criteria by means of which one can estimate the goodness of an opportunely trained neural network are also given (errors made vs. generalization capability). A great part of the adopted networks is based on adaptations of the Backpropagation algorithm; the other networks have been instead trained by means of algorithms based on statistical or geometric criteria. The Backpropagation algorithm has been improved by augmenting the degrees of freedom to the learning ability of a feed-forward neural network with the introduction of a spline adaptive activation function. A wide description has been given of the recurrent neural networks and particularly of the locally recurrent neural networks, networks for dynamic classification exploited in the automatic transcription of piano music. After a more or less rigorous definition of the concepts of classification and clustering, some paragraphs have been devoted to some statistical and geometric neural architectures, exploited in the implementation of static classifiers of common use and in particular in the application fields that have regarded my doctorate work. A separate paragraph has been devoted to the Simpson’s classifier and to the variants originated from my research work. They have revealed themselves to be static classifiers very simple to implement and at the same time very ductile and efficient, in many situations as well as regards the problem of musical source recognition. Two have been the choices in this case. In the first one, III these classifiers have been trained, by means of a pure supervised learning approach, while in the second the training algorithm, though keeping a substantially supervised nature, is prepared by a clustering phase, with the aim of improving, in terms of errors and generalization, the covering of the input space. Subsequently, the locally recurrent neural networks seen as dynamic classifiers are retrieved. However, their training has been rethought according to the effective reduction of the classification error instead of the classic mean-square error. The last three paragraphs have been devoted to a detailed description, in terms of specifications, implementative choices and final results, of the aforesaid fields of applications. The results obtained in all the three fields of application can be considered encouraging. Particularly, the recognition of musical instruments by means of the adopted neural networks has shown results tha can be considered out comparable if not better than those obtained by means of other techniques, but with considerably less complex structures. In case of the Automatic Transcription of piano pieces, the dynamic networks I adopted have given good results. Unfortunately, the required computational resources required by such networks cannot be considered negligible. As far as the medical applications, we are still in an incipent phase of the research. However, opinions expressed by those people who work in this field can be considered substantially eulogistic. The research activities my doctorate work is part of have been carried out in collaboration with the Department “INFOCOM” of the first University of Rome “La Sapienza”, as far as the recognition of musical instruments and the Automatic Transcription of piano pieces. The necessity to study the potentialities of neural classifiers in medical application has instead come from a profitable existing collaboration with the Istituto Superiore di Sanità in Rome.

碩士
中華大學
資訊工程學系碩士班
102
In this paper, wavelet transform based feature are extracted for automatic musical instrument recognition. The wavelet transform methods are proposed for feature extraction. First, one-dimension wavelet transform is used to decompose an imput signal into several subbands represented different frequency ranges. Then, the average energy and energy standard deviation are extracted from each subband as one-dimension wavelet transform features. The second method transforms the input music signal into a spectrogram and viewing it as a two-dimension image. Two-dimension wavelet transform is then used to decompose the spectrogram image into different subbands. The average energy and energy standard deviation are then extracted from each subbands and regarded as two-dimension wavelet transform features. Finally, nearest neighbor classification is used to recognize different musical instrumental sounds. The Euclidean distance is employed to calculate the distance between the testing data and each training data. In the experiments, the Iowa database and RWC database in which eight musical instruments including alto saxphone, bass, cello, flute, oboe, trumpet, viola, and violin are used to evaluated the recognition results. One-dimension wavelet transform features achieve the best recognition accuracy of 90% for Iowa database, and the best recognition accuracy of 83.16% for RWC database. In addition, We combine one-dimension wavelet transform features with two-dimension wavelet transform features. The best recognition accuracy of 98.31% for Iowa database, and the best recognition accuracy of 86.64% for RWC database.

碩士
中華大學
資訊工程學系碩士班
102
In this paper, modulation spectral analysis is applied for automatic musical instruments recognition. First we apply modulation spectral analysis on subband energy, subband average energy, subband energy standard deviation, extended subband energy, extended subband average energy, and extended subband energy standard deviation to extract variant modulation spectral features, Two-dimensional mel frequency cepstral coefficients is also employed to extract features from the musical signals. Linear discriminant analysis is exploited to increase the recognition accuracy at a lower dimensional feature vector space. Finally, the Eucidean distance is applied to estimate the distance between test note and each train note. In our experiments, eight instrumental classes including alto saxphone, bass, cello, flute, oboe, trumpet, violin, and viola were used to evaluate the performance. The proposed modulation spectral analysis of extended subband energy standard deviation feature achieve the highest recognition accuracy of 92.79%.

碩士
中華大學
資訊工程學系碩士班
103
In this paper, analysis of music signal using discrete wavelet packet transform (DWPT) is employed for musical instrument recognition. First, one-dimension DWPT is applied to decompose an input music signal into several subbands. Then, discriminative features were extracted from each subband. The features are average subband energy (SE), subband energy standard deviation (SSTD) and temporal subband average energy (TSE). Finally, we used the Euclidean distance to calculate the distance between the feature vector of testing data and each training data for musical instrument classification. In our experiments, eight kinds of musical instruments, including alto saxphone, bass, cello, flute, oboe, trumpet, viola, and violin, were used to evaluate the recognition rate. In the Iowa database, when using five level DWPT and combining SE, SSTD, and TSE, we can achieve the best recognition rate of 95.95%. In the RWC database, when using six level DWPT and combining SE, SSTD, and TSE, we can achieve the best recognition rate of 87.85%.

This thesis concerns with the recognition of music instruments from isolated notes. Music instrument recognition is a relatively nascent problem fast gaining importance not only because of the academic value the problem provides, but also for the potential it has in being able to realize applications like music content analysis, music transcription etc. Line spectral frequencies are proposed as features for music instrument recognition and shown to perform better than Mel filtered cepstral coefficients and linear prediction cepstral coefficients. Assuming a linear model of sound production, features based on the prediction residual, which represents the excitation signal, is proposed. Four improvements are proposed for classification using Gaussian mixture model (GMM) based classifiers. One of them involves characterizing the regions of overlap between classes in the feature space to improve classification. Applications to music instrument recognition and speaker recognition are shown. An experiment is proposed for discovering the hierarchy in music instrument in a data-driven manner. The hierarchy thus discovered closely corresponds to the hierarchy defined by musicians and experts and therefore shows that the feature space has successfully captured the required features for music instrument characterization.

This dissertation records the development of a process that enables within class classification of musical instruments. That is, a process that identifies a particular instrument of a given type - in this study four guitars and five violins. In recent years there have been numerous studies where between class classification has been attempted, but there have been no attempts at within class classification. Since timbre is the quality/quantity that enables one musical sound to be differentiated from another, before any classification can take place, a means to measure and describe it in physical terms needs to be devised. Towards this end, a review of musical timbre is presented which includes research into musical timbre from the work of Helmholtz through to the present. It also includes related work in speech recognition and musical instrument synthesis. The representation of timbre used in this study is influenced by the work of Hourdin and Charbonneau who used an adaption of multi-dimensional scaling, based on frequency analysis over time, to represent the evolution of each musical tone. A trajectory path, a plot of frequencies over time for each tone, was used to represent the evolution of each tone. This is achieved by taking a sequence of samples from the initial waveform and applying the discrete Fourier transform (DFT) or the constant Q transform (CQT) to achieve a frequency analysis of each data window. The classification technique used, is based on statistical distance methods. Two sets of data were recorded for each of the guitars and violins in the study across the pitch range of each instrument type. In the classi¯cation trials, one set of data was used as reference tones, and the other set, as test tones. To measure the similarity of timbre for a pair of tones, the closeness of the two trajectory paths was measured. This was achieved by summing the squared distances between corresponding points along the trajectory paths. With four guitars, a 97% correct classification rate was achieved for tones of the same pitch (fundamental frequency), and for five violins, a 94% correct classification rate was achieved for tones of the same pitch. The robustness of the classification system was tested by comparing a smaller portion of the whole tone, by comparing tones of differing pitch, and a number of other variations. It was found that classification of both guitars and violins was highly sensitive to pitch. The classification rate fell away markedly when tones of different pitch were compared. Further investigation was done to examine the timbre of each instrument across the range of the instrument. This conformed that the timbres of the guitar and violin are highly frequency dependent and suggested the presence of formants that is, certain fixed frequencies that are boosted when the tone contains harmonics at or near those frequencies.

This dissertation records the development of a process that enables within class classification of musical instruments. That is, a process that identifies a particular instrument of a given type - in this study four guitars and five violins. In recent years there have been numerous studies where between class classification has been attempted, but there have been no attempts at within class classification. Since timbre is the quality/quantity that enables one musical sound to be differentiated from another, before any classification can take place, a means to measure and describe it in physical terms needs to be devised. Towards this end, a review of musical timbre is presented which includes research into musical timbre from the work of Helmholtz through to the present. It also includes related work in speech recognition and musical instrument synthesis. The representation of timbre used in this study is influenced by the work of Hourdin and Charbonneau who used an adaption of multi-dimensional scaling, based on frequency analysis over time, to represent the evolution of each musical tone. A trajectory path, a plot of frequencies over time for each tone, was used to represent the evolution of each tone. This is achieved by taking a sequence of samples from the initial waveform and applying the discrete Fourier transform (DFT) or the constant Q transform (CQT) to achieve a frequency analysis of each data window. The classification technique used, is based on statistical distance methods. Two sets of data were recorded for each of the guitars and violins in the study across the pitch range of each instrument type. In the classi¯cation trials, one set of data was used as reference tones, and the other set, as test tones. To measure the similarity of timbre for a pair of tones, the closeness of the two trajectory paths was measured. This was achieved by summing the squared distances between corresponding points along the trajectory paths. With four guitars, a 97% correct classification rate was achieved for tones of the same pitch (fundamental frequency), and for five violins, a 94% correct classification rate was achieved for tones of the same pitch. The robustness of the classification system was tested by comparing a smaller portion of the whole tone, by comparing tones of differing pitch, and a number of other variations. It was found that classification of both guitars and violins was highly sensitive to pitch. The classification rate fell away markedly when tones of different pitch were compared. Further investigation was done to examine the timbre of each instrument across the range of the instrument. This conformed that the timbres of the guitar and violin are highly frequency dependent and suggested the presence of formants that is, certain fixed frequencies that are boosted when the tone contains harmonics at or near those frequencies.

Dissertação de Mestrado Integrado em Engenharia Electrotécnica e de Computadores apresentada à Faculdade de Ciências e Tecnologia
A identificação de instrumentos musicais continua um desafio por resolver na área de investigação em música, geralmente referida como Music Information Retreival (MIR). Este problema, fundamental para campos como a pesquisa por áudio, reconhecimento de género musical, recomendação de musica, ou a identificação de plágio, será abordado tendo em conta diversos métodos.A seguinte dissertação de mestrado apresenta um sistema de identificação de instrumentos que tem por base uma pequena parte da base de dados AudioSet com sons de instrumentos e que propõe o reconhecimento de áudio com base em imagens, neste caso espetogramas de mel, que representam o som que se pretende classificar.O OpenMic-2018 (OM18) é uma base de dados (BD) que surge no seguimento do AudioSet e com os mesmos ideais, mas direcionada para 20 classes de instrumentos musicais. Esta base de dados, publicada recentemente, conta ainda com poucos trabalhos que a abordem. Tentar-se-á superar os resultados já apresentados tanto através de abordagens originais como através de abordagens publicadas para o AudioSet. Trabalhos muito recentes utilizam modelos de atenção para classificar os exemplares do AudioSet e revelaram resultados muito positivos, pelo que também serão tidos em conta ao longo do projeto para a BD OM18.No âmbito do presente trabalho foi criada uma nova base de dados, \textbf{PureMic}, que tem por base as duas bases de dados já referenciadas. Esta é uma base de dados cujos exemplares são mais precisos e escolhidos de forma rigorosa, para poder contribuir para o classificador em tempo real e para uma melhoria das etiquetas do OM18, base de dados que ainda tem alguma falta de informação nesse aspeto.A seguinte dissertação faz então um resumo das abordagens a ser consideradas nomeadamente a implementação de redes neuronais convolucionais, muito utilizadas nesta área. Serão utilizadas as três bases de dados já referidas que providenciarão uma grande e suficiente quantidade de dados devidamente identificados.
Musical instruments recognition remains an unsolved challenge in Music Information Retreival (MIR). This problem, wich is fundamental for fields such as audio research, music genre recognition or music recommendation will be addressed with a variety of methods.This Master's dissertation presents an instrument identification system that is based on a small portion of AudioSet dataset with 20 musical instrument classes. This dataset proposes the recognition of audio events based on image inputs wich are log mel spectograms of sound events.OpenMic-2018 (OM18) is a dataset that extends the reach of AudioSet but targeted to only 20 classes of musical instruments. There are several publications arround AudioSet research. Since OpenMic its similar to AudioSet, some methods used in AudioSet will be aplied in OM18.In the context of this work, a new dataset was created, based on the two datasets already referenced. This is a dataset whose audio clips are more accurate and rigorously chosen to be able to contribute to the real-time classifier and to the improvement of OM18 labels.The following dissertation summarizes the approaches to be considered namely the implementation a convolutional neural netwrodk, widely used in this area. AudioSet, OpenMic-2018 and PureMic, will proveide a large and sufficiente amount of properly identified data. As AudioSet and OpenMic are Weakly Labeled Datasets (WLD), PureMic, a Strongly Labeled Dataset (SLD) will contribute to reduce the size of the other datasets but increase the quality of the labels.

博士
國立中興大學
電機工程學系所
101
This thesis applies wavelet features and soft computing-based classifiers to speech detection and musical instruments recognition problems. For speech detection, this thesis uses Haar wavelet energy and entropy (HWEE) as detection features. The Haar wavelet energy (HWE) is derived by using the robust band that shows the most significant difference between speech and nonspeech segments at different noise levels. Similarly, the wavelet energy entropy (WEE) is computed by selecting the two wavelet energy bands of which entropy shows the most significant speech/nonspeech difference. The HWEE features are fed as inputs to a recurrent self-evolving interval type-2 fuzzy neural network (RSEIT2FNN) for classification. The RSEIT2FNN is used because it uses type-2 fuzzy sets, which are more robust to noise than type-1 fuzzy sets. The recurrent structure in the RSEIT2FNN helps to remember the context information of a test frame. The HWEE-based RSEIT2FNN detection was applied to speech detection in different noisy environments with different noise levels. The most critical technology of musical instrument recognition is the extraction of musical instrument features, especially those distinguishing ones. To this end, this thesis proposes the incorporation of HWE and wavelet packet decomposition energy features into traditional audio features to improve recognition rate. For classifier design, this thesis proposes a divide-and-conquer support vector machine (SVM) classification technique. This technique efficiently captures the context information in audio signals with high-dimensional features for recognition improvement. Experimental results on the recognition of ten musical instruments show the effectiveness of the proposed recognition approach.