The aim of this thesis is to review the current state of machine learning in music composition and to train a computer on Beatles' songs using research project Magenta from the Google Brain Team to produce its own music. In order to explore the qualities of the generated music more thoroughly, we restrict our- selves to monophonic melodies only. We train three deep learning models with three different configurations (Basic, Lookback, and Attention) and compare generated results. Even though the generated music is not as interesting as the original Beatles, it is quite likable. According to our analysis based on musically informed metrics, generated melodies differ from the original ones especially in lengths of notes and in pitch differences between consecutive notes. Generated melodies tend to use shorter notes and higher pitch differences. In theoretical background, we cover the most commonly used machine learning algorithms, introduce neural networks and review related work of music generation. 1

Semantic parsing is the process of mapping a natural-language sentence into a machine-readable, formal representation of its meaning. The LSTM Encoder-Decoder is a neural architecture with the ability to map a source language into a target one. We are interested in the problem of mapping natural language into SPARQL queries, and we seek to contribute with strategies that do not rely on handcrafted rules, high-quality lexicons, manually-built templates or other handmade complex structures. In this context, we present two contributions to the problem of semantic parsing departing from the LSTM encoder-decoder. While natural language has well defined vector representation methods that use a very large volume of texts, formal languages, like SPARQL queries, suffer from lack of suitable methods for vector representation. In the first contribution we improve the representation of SPARQL vectors. We start by obtaining an alignment matrix between the two vocabularies, natural language and SPARQL terms, which allows us to refine a vectorial representation of SPARQL items. With this refinement we obtained better results in the posterior training for the semantic parsing model. In the second contribution we propose a neural architecture, that we call Encoder CFG-Decoder, whose output conforms to a given context-free grammar. Unlike the traditional LSTM encoder-decoder, our model provides a grammatical guarantee for the mapping process, which is particularly important for practical cases where grammatical errors can cause critical failures. Results confirm that any output generated by our model obeys the given CFG, and we observe a translation accuracy improvement when compared with other results from the literature. / A análise semântica é o processo de mapear uma sentença em linguagem natural para uma representação formal, interpretável por máquina, do seu significado. O LSTM Encoder-Decoder é uma arquitetura de rede neural com a capacidade de mapear uma sequência de origem para uma sequência de destino. Estamos interessados no problema de mapear a linguagem natural em consultas SPARQL e procuramos contribuir com estratégias que não dependam de regras artesanais, léxico de alta qualidade, modelos construídos manualmente ou outras estruturas complexas feitas à mão. Neste contexto, apresentamos duas contribuições para o problema de análise semântica partindo da arquitetura LSTM Encoder-Decoder. Enquanto para a linguagem natural existem métodos de representação vetorial bem definidos que usam um volume muito grande de textos, as linguagens formais, como as consultas SPARQL, sofrem com a falta de métodos adequados para representação vetorial. Na primeira contribuição, melhoramos a representação dos vetores SPARQL. Começamos obtendo uma matriz de alinhamento entre os dois vocabulários, linguagem natural e termos SPARQL, o que nos permite refinar uma representação vetorial dos termos SPARQL. Com esse refinamento, obtivemos melhores resultados no treinamento posterior para o modelo de análise semântica. Na segunda contribuição, propomos uma arquitetura neural, que chamamos de Encoder CFG-Decoder, cuja saída está de acordo com uma determinada gramática livre de contexto. Ao contrário do modelo tradicional LSTM Encoder-Decoder, nosso modelo fornece uma garantia gramatical para o processo de mapeamento, o que é particularmente importante para casos práticos nos quais erros gramaticais podem causar falhas críticas em um compilador ou interpretador. Os resultados confirmam que qualquer resultado gerado pelo nosso modelo obedece à CFG dada, e observamos uma melhora na precisão da tradução quando comparada com outros resultados da literatura.

Un robot d'assistance sociale (SAR) est destiné à engager les gens dans une interaction située comme la surveillance de l'exercice physique, la réadaptation neuropsychologique ou l'entraînement cognitif. Alors que les comportements interactifs de ces systèmes sont généralement scriptés, nous discutons ici du cadre d’apprentissage de comportements interactifs multimodaux qui est proposé par le projet SOMBRERO.Dans notre travail, nous avons utilisé l'apprentissage par démonstration afin de fournir au robot des compétences nécessaires pour effectuer des tâches collaboratives avec des partenaires humains. Il y a trois étapes principales d'apprentissage de l'interaction par démonstration: (1) recueillir des comportements interactifs représentatifs démontrés par des tuteurs humains; (2) construire des modèles des comportements observés tout en tenant compte des connaissances a priori (modèle de tâche et d'utilisateur, etc.); et ensuite (3) fournir au robot-cible des contrôleurs de gestes appropriés pour exécuter les comportements souhaités.Les modèles multimodaux HRI (Human-Robot Interaction) sont fortement inspirés des interactions humain-humain (HHI). Le transfert des comportements HHI aux modèles HRI se heurte à plusieurs problèmes: (1) adapter les comportements humains aux capacités interactives du robot en ce qui concerne ses limitations physiques et ses capacités de perception, d'action et de raisonnement limitées; (2) les changements drastiques des comportements des partenaires humains face aux robots ou aux agents virtuels; (3) la modélisation des comportements interactifs conjoints; (4) la validation des comportements robotiques par les partenaires humains jusqu'à ce qu'ils soient perçus comme adéquats et significatifs.Dans cette thèse, nous étudions et faisons des progrès sur ces quatre défis. En particulier, nous traitons les deux premiers problèmes (transfert de HHI vers HRI) en adaptant le scénario et en utilisant la téléopération immersive. En outre, nous utilisons des réseaux neuronaux récurrents pour modéliser les comportements interactifs multimodaux (tels que le discours, le regard, les mouvements de bras, les mouvements de la tête, les canaux). Ces techniques récentes surpassent les méthodes traditionnelles (Hidden Markov Model, Dynamic Bayesian Network, etc.) en termes de précision et de coordination inter-modalités. A la fin de cette thèse, nous évaluons une première version de robot autonome équipé des modèles construits par apprentissage. / A socially assistive robot (SAR) is meant to engage people into situated interaction such as monitoring physical exercise, neuropsychological rehabilitation or cognitive training. While the interactive behavioral policies of such systems are mainly hand-scripted, we discuss here key features of the training of multimodal interactive behaviors in the framework of the SOMBRERO project.In our work, we used learning by demonstration in order to provide the robot with adequate skills for performing collaborative tasks in human centered environments. There are three main steps of learning interaction by demonstration: we should (1) collect representative interactive behaviors from human coaches; (2) build comprehensive models of these overt behaviors while taking into account a priori knowledge (task and user model, etc.); and then (3) provide the target robot with appropriate gesture controllers to execute the desired behaviors.Multimodal HRI (Human-Robot Interaction) models are mostly inspired by Human-Human interaction (HHI) behaviors. Transferring HHI behaviors to HRI models faces several issues: (1) adapting the human behaviors to the robot’s interactive capabilities with regards to its physical limitations and impoverished perception, action and reasoning capabilities; (2) the drastic changes of human partner behaviors in front of robots or virtual agents; (3) the modeling of joint interactive behaviors; (4) the validation of the robotic behaviors by human partners until they are perceived as adequate and meaningful.In this thesis, we study and make progress over those four challenges. In particular, we solve the two first issues (transfer from HHI to HRI) by adapting the scenario and using immersive teleoperation. In addition, we use Recurrent Neural Networks to model multimodal interactive behaviors (such as speech, gaze, arm movements, head motion, backchannels) that surpass traditional methods (Hidden Markov Model, Dynamic Bayesian Network, etc.) in both accuracy and coordination between the modalities. We also build and evaluate a proof-of-concept autonomous robot to perform the tasks.

This thesis deals with the possibility of using neural networks in the analysis of a racing vehicle telemetry data. The analysis consists of finding and classifying the driving states of the vehicle as well as reconstruction and prediction of the data signal. The thesis consists of a research of machine learning focusing on neural networks. It describes the main types of the most commonly used neural networks with the description of their structure and their learning process. It also deals with the problem of classification and an in-depth analysis of the signal prediction and reconstruction. The conclusion is devoted to the implementation of the network into a TeleMatrix application.

La tâche de segmentation et de regroupement en locuteurs (speaker diarization) consiste à identifier "qui parle quand" dans un flux audio sans connaissance a priori du nombre de locuteurs ou de leur temps de parole respectifs. Les systèmes de segmentation et de regroupement en locuteurs sont généralement construits en combinant quatre étapes principales. Premièrement, les régions ne contenant pas de parole telles que les silences, la musique et le bruit sont supprimées par la détection d'activité vocale (VAD). Ensuite, les régions de parole sont divisées en segments homogènes en locuteur par détection des changements de locuteurs, puis regroupées en fonction de l'identité du locuteur. Enfin, les frontières des tours de parole et leurs étiquettes sont affinées avec une étape de re-segmentation. Dans cette thèse, nous proposons d'aborder ces quatre étapes avec des approches fondées sur les réseaux de neurones. Nous formulons d’abord le problème de la segmentation initiale (détection de l’activité vocale et des changements entre locuteurs) et de la re-segmentation finale sous la forme d’un ensemble de problèmes d’étiquetage de séquence, puis nous les résolvons avec des réseaux neuronaux récurrents de type Bi-LSTM (Bidirectional Long Short-Term Memory). Au stade du regroupement des régions de parole, nous proposons d’utiliser l'algorithme de propagation d'affinité à partir de plongements neuronaux de ces tours de parole dans l'espace vectoriel des locuteurs. Des expériences sur un jeu de données télévisées montrent que le regroupement par propagation d'affinité est plus approprié que le regroupement hiérarchique agglomératif lorsqu'il est appliqué à des plongements neuronaux de locuteurs. La segmentation basée sur les réseaux récurrents et la propagation d'affinité sont également combinées et optimisées conjointement pour former une chaîne de regroupement en locuteurs. Comparé à un système dont les modules sont optimisés indépendamment, la nouvelle chaîne de traitements apporte une amélioration significative. De plus, nous proposons d’améliorer l'estimation de la matrice de similarité par des réseaux neuronaux récurrents, puis d’appliquer un partitionnement spectral à partir de cette matrice de similarité améliorée. Le système proposé atteint des performances à l'état de l'art sur la base de données de conversation téléphonique CALLHOME. Enfin, nous formulons le regroupement des tours de parole en mode séquentiel sous la forme d'une tâche supervisée d’étiquetage de séquence et abordons ce problème avec des réseaux récurrents empilés. Pour mieux comprendre le comportement du système, une analyse basée sur une architecture de codeur-décodeur est proposée. Sur des exemples synthétiques, nos systèmes apportent une amélioration significative par rapport aux méthodes de regroupement traditionnelles. / Speaker diarization is the task of determining "who speaks when" in an audio stream that usually contains an unknown amount of speech from an unknown number of speakers. Speaker diarization systems are usually built as the combination of four main stages. First, non-speech regions such as silence, music, and noise are removed by Voice Activity Detection (VAD). Next, speech regions are split into speaker-homogeneous segments by Speaker Change Detection (SCD), later grouped according to the identity of the speaker thanks to unsupervised clustering approaches. Finally, speech turn boundaries and labels are (optionally) refined with a re-segmentation stage. In this thesis, we propose to address these four stages with neural network approaches. We first formulate both the initial segmentation (voice activity detection and speaker change detection) and the final re-segmentation as a set of sequence labeling problems and then address them with Bidirectional Long Short-Term Memory (Bi-LSTM) networks. In the speech turn clustering stage, we propose to use affinity propagation on top of neural speaker embeddings. Experiments on a broadcast TV dataset show that affinity propagation clustering is more suitable than hierarchical agglomerative clustering when applied to neural speaker embeddings. The LSTM-based segmentation and affinity propagation clustering are also combined and jointly optimized to form a speaker diarization pipeline. Compared to the pipeline with independently optimized modules, the new pipeline brings a significant improvement. In addition, we propose to improve the similarity matrix by bidirectional LSTM and then apply spectral clustering on top of the improved similarity matrix. The proposed system achieves state-of-the-art performance in the CALLHOME telephone conversation dataset. Finally, we formulate sequential clustering as a supervised sequence labeling task and address it with stacked RNNs. To better understand its behavior, the analysis is based on a proposed encoder-decoder architecture. Our proposed systems bring a significant improvement compared with traditional clustering methods on toy examples.

Les radiologues utilisent au quotidien des solutions d'imagerie médicale pour le diagnostic. L'amélioration de l'expérience utilisateur est toujours un axe majeur de l'effort continu visant à améliorer la qualité globale et l'ergonomie des produits logiciels. Les applications de monitoring permettent en particulier d'enregistrer les actions successives effectuées par les utilisateurs dans l'interface du logiciel. Ces interactions peuvent être représentées sous forme de séquences d'actions. Sur la base de ces données, ce travail traite de deux sujets industriels : les pannes logicielles et l'ergonomie des logiciels. Ces deux thèmes impliquent d'une part la compréhension des modes d'utilisation, et d'autre part le développement d'outils de prédiction permettant soit d'anticiper les pannes, soit d'adapter dynamiquement l'interface logicielle en fonction des besoins des utilisateurs. Tout d'abord, nous visons à identifier les origines des crashes du logiciel qui sont essentielles afin de pouvoir les corriger. Pour ce faire, nous proposons d'utiliser un test binomial afin de déterminer quel type de pattern est le plus approprié pour représenter les signatures de crash. L'amélioration de l'expérience utilisateur par la personnalisation et l'adaptation des systèmes aux besoins spécifiques de l'utilisateur exige une très bonne connaissance de la façon dont les utilisateurs utilisent le logiciel. Afin de mettre en évidence les tendances d'utilisation, nous proposons de regrouper les sessions similaires. Nous comparons trois types de représentation de session dans différents algorithmes de clustering. La deuxième contribution de cette thèse concerne le suivi dynamique de l'utilisation du logiciel. Nous proposons deux méthodes -- basées sur des représentations différentes des actions d'entrée -- pour répondre à deux problématiques industrielles distinctes : la prédiction de la prochaine action et la détection du risque de crash logiciel. Les deux méthodologies tirent parti de la structure récurrente des réseaux LSTM pour capturer les dépendances entre nos données séquentielles ainsi que leur capacité à traiter potentiellement différents types de représentations d'entrée pour les mêmes données. / Radiologists use medical imaging solutions on a daily basis for diagnosis. Improving user experience is a major line of the continuous effort to enhance the global quality and usability of software products. Monitoring applications enable to record the evolution of various software and system parameters during their use and in particular the successive actions performed by the users in the software interface. These interactions may be represented as sequences of actions. Based on this data, this work deals with two industrial topics: software crashes and software usability. Both topics imply on one hand understanding the patterns of use, and on the other developing prediction tools either to anticipate crashes or to dynamically adapt software interface according to users' needs. First, we aim at identifying crash root causes. It is essential in order to fix the original defects. For this purpose, we propose to use a binomial test to determine which type of patterns is the most appropriate to represent crash signatures. The improvement of software usability through customization and adaptation of systems to each user's specific needs requires a very good knowledge of how users use the software. In order to highlight the trends of use, we propose to group similar sessions into clusters. We compare 3 session representations as inputs of different clustering algorithms. The second contribution of our thesis concerns the dynamical monitoring of software use. We propose two methods -- based on different representations of input actions -- to address two distinct industrial issues: next action prediction and software crash risk detection. Both methodologies take advantage of the recurrent structure of LSTM neural networks to capture dependencies among our sequential data as well as their capacity to potentially handle different types of input representations for the same data.

LSTM neural networks have been used for speech recognition, image recognition and other artificial intelligence applications for many years. Most applications perform the LSTM algorithm and the required calculations on cloud computers. Off-line solutions include the use of FPGAs and GPUs but the most promising solutions include ASIC accelerators designed for this purpose only. This report presents an ASIC design capable of performing the multiple iterations of the LSTM algorithm on a unidirectional and without peepholes neural network architecture. The proposed design provides arithmetic level parallelism options as blocks are instantiated based on parameters. The internal structure of the design implements pipelined, parallel or serial solutions depending on which is optimal in every case. The implications concerning these decisions are discussed in detail in the report. The design process is described in detail and the evaluation of the design is also presented to measure accuracy and error of the design output.This thesis work resulted in a complete synthesizable ASIC design implementing an LSTM layer, a Fully Connected layer and a Softmax layer which can perform classification of data based on trained weight matrices and bias vectors. The design primarily uses 16-bit fixed point format with 5 integer and 11 fractional bits but increased precision representations are used in some blocks to reduce error output. Additionally, a verification environment has also been designed and is capable of performing simulations, evaluating the design output by comparing it with results produced from performing the same operations with 64-bit floating point precision on a SystemVerilog testbench and measuring the encountered error. The results concerning the accuracy and the design output error margin are presented in this thesis report. The design went through Logic and Physical synthesis and successfully resulted in a functional netlist for every tested configuration. Timing, area and power measurements on the generated netlists of various configurations of the design show consistency and are reported in this report. / LSTM neurala nätverk har använts för taligenkänning, bildigenkänning och andra artificiella intelligensapplikationer i många år. De flesta applikationer utför LSTM-algoritmen och de nödvändiga beräkningarna i digitala moln. Offline lösningar inkluderar användningen av FPGA och GPU men de mest lovande lösningarna inkluderar ASIC-acceleratorer utformade för endast dettaändamål. Denna rapport presenterar en ASIC-design som kan utföra multipla iterationer av LSTM-algoritmen på en enkelriktad neural nätverksarkitetur utan peepholes. Den föreslagna designed ger aritmetrisk nivå-parallellismalternativ som block som är instansierat baserat på parametrar. Designens inre konstruktion implementerar pipelinerade, parallella, eller seriella lösningar beroende på vilket anternativ som är optimalt till alla fall. Konsekvenserna för dessa beslut diskuteras i detalj i rapporten. Designprocessen beskrivs i detalj och utvärderingen av designen presenteras också för att mäta noggrannheten och felmarginal i designutgången. Resultatet av arbetet från denna rapport är en fullständig syntetiserbar ASIC design som har implementerat ett LSTM-lager, ett fullständigt anslutet lager och ett Softmax-lager som kan utföra klassificering av data baserat på tränade viktmatriser och biasvektorer. Designen använder huvudsakligen 16bitars fast flytpunktsformat med 5 heltal och 11 fraktions bitar men ökade precisionsrepresentationer används i vissa block för att minska felmarginal. Till detta har även en verifieringsmiljö utformats som kan utföra simuleringar, utvärdera designresultatet genom att jämföra det med resultatet som produceras från att utföra samma operationer med 64-bitars flytpunktsprecision på en SystemVerilog testbänk och mäta uppstådda felmarginal. Resultaten avseende noggrannheten och designutgångens felmarginal presenteras i denna rapport.Designen gick genom Logisk och Fysisk syntes och framgångsrikt resulterade i en funktionell nätlista för varje testad konfiguration. Timing, area och effektmätningar på den genererade nätlistorna av olika konfigurationer av designen visar konsistens och rapporteras i denna rapport.

The music industry has seen a great influx of new channels to browse and distribute music. This does not come without drawbacks. As the data rapidly increases, manual curation becomes a much more difficult task. Audio files have a plethora of features that could be used to make parts of this process a lot easier. It is possible to extract these features, but the best way to handle these for different tasks is not always known. This thesis compares the two deep learning models, convolutional neural network (CNN) and long short-term memory (LSTM), for music genre classification when trained using mel-frequency cepstral coefficients (MFCCs) in hopes of making audio data as useful as possible for future usage. These models were tested on two different datasets, GTZAN and FMA, and the results show that the CNN had a 56.0% and 50.5% prediction accuracy, respectively. This outperformed the LSTM model that instead achieved a 42.0% and 33.5% prediction accuracy. / Musikindustrin har sett en stor ökning i antalet sätt att hitta och distribuera musik. Det kommer däremot med sina nackdelar, då mängden data ökar fort så blir det svårare att hantera den på ett bra sätt. Ljudfiler har mängder av information man kan extrahera och därmed göra den här processen enklare. Det är möjligt att använda sig av de olika typer av information som finns i filen, men bästa sättet att hantera dessa är inte alltid känt. Den här rapporten jämför två olika djupinlärningsmetoder, convolutional neural network (CNN) och long short-term memory (LSTM), tränade med mel-frequency cepstral coefficients (MFCCs) för klassificering av musikgenre i hopp om att göra ljuddata lättare att hantera inför framtida användning. Modellerna testades på två olika dataset, GTZAN och FMA, där resultaten visade att CNN:et fick en träffsäkerhet på 56.0% och 50.5% tränat på respektive dataset. Denna utpresterade LSTM modellen som istället uppnådde en träffsäkerhet på 42.0% och 33.5%.

Artificial intelligence is receiving a great deal of attention in various fields of science and engineering due to its promising applications. In today’s society, weather classification models with high accuracy are of utmost importance. An alternative to using conventional weather radars is to use measured attenuation data in microwave links as the input to deep learning-based weather classification models. Detecting anomalies in the measured attenuation data is of great importance as the output of a classification model cannot be trusted if the input to the classification model contains anomalies. Designing an accurate classification model poses some challenges due to the absence of predefined features to discriminate among the various weather conditions, and due to specific domain requirements in terms of execution time and detection sensitivity. In this thesis we investigate the relationship between anomalies in signal attenuation data, which is the input to a weather classification model, and the model’s misclassifications. To this end, we propose and evaluate two deep learning models based on long short-term memory networks (LSTM) and convolutional neural networks (CNN) for anomaly detection in a weather classification problem. We evaluate the feasibility and possible generalizations of the proposed methodology in an industrial case study at Ericsson AB, Sweden. The results show that both proposed methods can detect anomalies that correlate with misclassifications made by the weather classifier. Although the LSTM performed better than the CNN with regards to top performance on one link and average performance across all 5 tested links, the CNN performance is shown to be more consistent. / Artificiell intelligens har fått mycket uppmärksamhet inom olika teknik- och vetenskapsområden på grund av dess många lovande tillämpningar. I dagens samhälle är väderklassificeringsmodeller med hög noggrannhet av yttersta vikt. Ett alternativ till att använda konventionell väderradar är att använda uppmätta dämpningsdata i mikrovågslänkar som indata till djupinlärningsbaserade väderklassificeringsmodeller. Detektering av avvikelser i uppmätta dämpningsdata är av stor betydelse eftersom en klassificeringsmodells pålitlighet minskar om träningsdatat innehåller avvikelser. Att utforma en noggrann klassificeringsmodell är svårt på grund av bristen på fördefinierade kännetecken för olika typer av väderförhållanden, och på grund av de specifika domänkrav som ofta ställs när det gäller exekveringstid och detekteringskänslighet. I det här examensarbetet undersöker vi förhållandet mellan avvikelser i uppmätta dämpningsdata från mikrovågslänkar, och felklassificeringar gjorda av en väderklassificeringsmodell. För detta ändamål utvärderar vi avvikelsedetektering inom ramen för väderklassificering med hjälp av två djupinlärningsmodeller, baserade på long short-term memory-nätverk (LSTM) och faltningsnätverk (CNN). Vi utvärderar genomförbarhet och generaliserbarhet av den föreslagna metodiken i en industriell fallstudie hos Ericsson AB. Resultaten visar att båda föreslagna metoder kan upptäcka avvikelser som korrelerar med felklassificeringar gjorda av väderklassificeringsmodellen. LSTM-modellen presterade bättre än CNN-modellen både med hänsyn till toppprestanda på en länk och med hänsyn till genomsnittlig prestanda över alla 5 testade länkar, men CNNmodellens prestanda var mer konsistent.

Traditionally, in Speech Emotion Recognition, models require a large number of manually engineered features and intermediate representations such as spectrograms for training. However, to hand-engineer such features often requires both expert domain knowledge and resources. Recently, with the emerging paradigm of deep-learning, end-to-end models that extract features themselves and learn from the raw speech signal directly have been explored. A previous approach has been to combine multiple parallel CNNs with different filter lengths to extract multiple temporal features from the audio signal, and then feed the resulting sequence to a recurrent block. Also, other recent work present high accuracies when utilizing local feature learning blocks (LFLBs) for reducing the dimensionality of a raw audio signal, extracting the most important information. Thus, this study will combine the idea of LFLBs for feature extraction with a block of parallel CNNs with different filter lengths for capturing multitemporal features; this will finally be fed into an LSTM layer for global contextual feature learning. To the best of our knowledge, such a combined architecture has yet not been properly investigated. Further, this study will investigate different configurations of such an architecture. The proposed model is then trained and evaluated on the well-known speech databases EmoDB and RAVDESS, both in a speaker-dependent and speaker-independent manner. The results indicate that the proposed architecture can produce comparable results with state-of-the-art; despite excluding data augmentation and advanced pre-processing. It was reported 3 parallel CNN pipes yielded the highest accuracy, together with a series of modified LFLBs that utilize averagepooling and ReLU activation. This shows the power of leaving the feature learning up to the network and opens up for interesting future research on time-complexity and trade-off between introducing complexity in pre-processing or in the model architecture itself. / Traditionellt sätt, vid talbaserad känsloigenkänning, kräver modeller ett stort antal manuellt konstruerade attribut och mellanliggande representationer, såsom spektrogram, för träning. Men att konstruera sådana attribut för hand kräver ofta både domänspecifika expertkunskaper och resurser. Nyligen har djupinlärningens framväxande end-to-end modeller, som utvinner attribut och lär sig direkt från den råa ljudsignalen, undersökts. Ett tidigare tillvägagångssätt har varit att kombinera parallella CNN:er med olika filterlängder för att extrahera flera temporala attribut från ljudsignalen och sedan låta den resulterande sekvensen passera vidare in i ett så kallat Recurrent Neural Network. Andra tidigare studier har också nått en hög noggrannhet när man använder lokala inlärningsblock (LFLB) för att reducera dimensionaliteten hos den råa ljudsignalen, och på så sätt extraheras den viktigaste informationen från ljudet. Således kombinerar denna studie idén om att nyttja LFLB:er för extraktion av attribut, tillsammans med ett block av parallella CNN:er som har olika filterlängder för att fånga multitemporala attribut; detta kommer slutligen att matas in i ett LSTM-lager för global inlärning av kontextuell information. Så vitt vi vet har en sådan kombinerad arkitektur ännu inte undersökts. Vidare kommer denna studie att undersöka olika konfigurationer av en sådan arkitektur. Den föreslagna modellen tränas och utvärderas sedan på de välkända taldatabaserna EmoDB och RAVDESS, både via ett talarberoende och talaroberoende tillvägagångssätt. Resultaten indikerar att den föreslagna arkitekturen kan ge jämförbara resultat med state-of-the-art, trots att ingen ökning av data eller avancerad förbehandling har inkluderats. Det rapporteras att 3 parallella CNN-lager gav högsta noggrannhet, tillsammans med en serie av modifierade LFLB:er som nyttjar average-pooling och ReLU som aktiveringsfunktion. Detta visar fördelarna med att lämna inlärningen av attribut till nätverket och öppnar upp för intressant framtida forskning kring tidskomplexitet och avvägning mellan introduktion av komplexitet i förbehandlingen eller i själva modellarkitekturen.