theses

This work proposes a real-time, small-footprint emergency siren detection system based on deep learning, designed to operate efficiently inside car cabin environments while meeting strict latency and computational constraints.

This work proposes a data-driven and spectrum-based BWE method that leverages the source-filter representation of speech. The model predicts an amplitude mask applied to a shaped noise signal. Results demonstrated that the proposed approach achieves competitive perceptual quality, while drastically reducing computational complexity.

Input gain variations can significantly impact the performance of DNN-based real-time speech processing systems. This thesis explores three methods to enhance robustness against these variations: Gain-Augmented Training, Differential Features, and Smoothed Frame Normalization. Experimental results show that these approaches improve the consistency and reliability of DNN outputs under varying input gain condition.

This work tackles key challenges in Speaker Verification (SV) by introducing a novel, lightweight SV system designed for real-time applications in noisy and reverberant environments. The system leverages advanced convolutional techniques within a Deep Neural Network (DNN) and real-time pooling layers to enhance responsiveness and stability across various acoustic conditions. While it may not achieve the highest performance levels, it excels in real-time processing, making it ideal for dynamic environments where speed and computational efficiency are crucial.

We propose a cascaded network with a lightweight phase-unaware approach and an optional more computationally demanding phase-aware stage to perform single-channel Speech Enahncement based on Deep Learning (DL). Our solution performs as good as more complex baselines in terms of parameters and Floating Point Operations (FLOPs) according to both objective quality metrics and subjective evaluations

We propose an innovative technique based on the use of a Convolutional Neural Network (CNN), designed to offer a small-footprint and optimized computational performance, for systems that workin real-time, with minimal latency.

We address the multichannel speaker separation problem, and we propose two causal and lightweight Deep Neural Network (DNN) models that can adapt to a wide range of microphone positions and distances. The problem focuses on the automotive scenario.

In this work, we propose a speech enhancement solution based on Deep Neural Networks that withstands the strict requirements imposed by embedded devices in terms of memory footprint and processing power. The proposed approach operates in real-time, extracting perceptually-relevant features in an efficient fashion.

A novel approach, using a talk state detector (TSD) to enhance the performance of a linear acoustic echo cancellation. It consists of a fully convolutional neural network classifier that performs causal processing to meet the real-time requiremment with less than 8,000 trainable parameters.

A speaker recognition system is a technology that aims to recognize a person’s identity based on their voice. In this thesis, we propose a low-latency speaker recognition system based on Deep Neural Networks.

We propose a hybrid beamforming solution that combines two methods: one that is efficient for signals with high input SNR and one with low input SNR. Results show that our SCM-based hybrid solution outperforms most SCM-based methods and exhibits a lower computational complexity.

Techniques employed to detect the presence or absence of human voice in an audio signal are called Voice Activity Detection (VAD) algorithms. Our approach optimizes both the feature extraction and the classification performed by the deep neural network. The goal is to comply with requirements imposed by embedded systems.

We present a hybrid model for personalized similarity modeling that relies on both content-based and user-related similarity information. We exploit a non-metric scaling technique to first elaborate a low-dimensional space (or embedding) which fulfills the similarity information provided by the user, and a regression technique to learn a mapping between content-based information and embedding-related information.

In this work, we propose an approach to apply transfer learning for beat tracking. We use a deep RNN as the starting network trained on popular music, and we transfer it to track beats of folk music. Moreover, we test if the resultant models are able to deal with highly variable music, such as Greek folk music.

In this thesis we propose an approach to model a personalized music similarity metric based on a Deep Neural Network. We use a first stage for learning a generic music similarity metric relying on a great amount of data, and a second stage for customizing it using personalized annotations collected through a survey.

In this study we propose an automatic playlist generation approach which analyzes hand-crafted playlists, understands their structure and generates new playlists accordingly. We have adopted a deep learning architecture, in particular a Recurrent Neural Network, which is specialized in sequence modeling.

Analysis algorithm where we use a Deep Belief Network to extract a sequence of descriptors that is successively given as input to several Music Structural Analysis algorithms presented in literature.

In this work we propose an approach for music high-level description and music retrieval, that we named Contextual-related semantic model. Our method defines different semantic contexts and dimensional semantic relations between music descriptors belonging to the same context.