8 research outputs found
Deep learning for speech separation.
No full textSpeech source separation aims to estimate one or more individual sources from mixtures of multiple sound sources, e.g. speech, noise and music. While humans have an innate ability to separate sources in a sound mixture, this is not a trivial task for computers. In this thesis, we study the problem of speech separation, with a varying degree of complexity with respect to room reverberation, the number of speech sources and the number of microphones available for capturing the sources. We focus on the stateof- the-art deep learning techniques, and investigate the problem of separating speech sources from binaural and B-format mixtures obtained in real reverberant rooms. First, we evaluate a baseline system for binaural speech separation, where fullyconnected Deep Neural Networks (DNNs) and spatial features, such as Interaural Level Difference (ILD) and Interaural Phase Difference (IPD), are used. We further extend this baseline by using the dropout technique to mitigate the overfitting problem and adding spectral features, such as the Log-Power Spectrogram (LPS), to improve the separation performance. Second, we develop a Convolutional Neural Networks (CNNs)-based binaural speech separation system. We then study the potential of using data augmentation techniques to improve speech separation quality. In particular, we introduce contextual frames expansion, by including the information from neighbouring time frames, before and after a given time frame. Finally, we study the use of deep learning methods for B-format recordings. This allows the pressure gradient information to be exploited, in addition to the widely used acoustic pressure information, for deriving the angular features for source separation. Extensive experiments have been performed on two datasets captured in five different rooms in the University of Surrey. The proposed methods are shown to offer improved performance over the state-of-the-art, in terms of separation quality and intelligibility
Characterization and simulation of drift chambers prototypes for the MEG II experiment
No full textThe MEG experiment, located at the Paul Scherrer Institut near Zurich (Switzerland),
searches for the lepton flavour violating decay
μ+ → e+ + γ
(1)
The MEG collaboration has recently published a new upper limit on the branching ratio
BR(μ+ → e+ + γ) < 5.7 · 10−13 (90% C.L.)
The Standard Model does not predict violation of the flavour of charged-leptons, that
means the decay (1) is forbidden by this theory, unless the recently discovered neu-
trino oscillations are incorporated in the Standard Model: in this case, an experimentally
unattainable branching ratio ∼ 10−55 is expected. Some Super-Symmetric extensions to
the Standard Model predict the decay (1) in the range 10−11 − 10−14 , making the search
worthwhile with a new upgraded apparatus.
The MEG II experiment aims to reach a ten times lower limit on the branching ratio
BR(μ+ → e+ + γ) < 6.0 · 10−14
in 3 years of data taking; studies on the upgrade have already started and will continue
until 2015, then a first test run is expected to be done at the end of 2015.
In order to achieve this result, an improvement of the detectors resolutions is mandatory:
the two main detectors involved in the upgrade process are the magnetic spectrometer
and the LXe calorimeter which, respectively, measure momentum and emission point of
the positron and the photon energy, as well as the position of the photon conversion.
The new improved drift chamber is designed with more than a thousand wires in a stereo
configuration, immersed in a helium-isobutane gas mixture as active medium. The cham-
ber allows for a 3-dimensional reconstruction of the positron trajectory.
An increase by a factor 3 of the MEG muon-stopping rate is also foreseen in the MEG II
upgrade, but it will affect the drift chamber performances during the detector operation.
Fragmentation of the gas molecules in the avalanches causes formation of deposits along
the wires, with a reduction of the gas gain or, in the extreme case, continous discharge of
the chamber: this phenomenon is called ‘ageing’. The collected charge density on wires is
of the order of a few tenth of Coulombs per centimetre during the whole experiment life-
time, so a study of how much the chamber performances change due to ageing is crucial.
This thesis analyzes different aspects that must be taken into account in the construction
of the final MEG II drift chamber, through realization and characterization of several
different prototypes at the Pisa INFN laboratories.
The ageing effects have been studied in a series of prototypes, based on a single drift cell
configuration, some of them built using a preliminary set of wires and one with the final
chamber wires; also the insertion of contaminant material inside the gas volume has been
studied.
Inspections of the aged wires with the SEM (Scanning Electron Microscope) are presentedalong with an EDX (Energy Dispersive X-ray spectroscopy) analysis.
The main feature of a drift chamber is the capability of tracking a charged particle po-
sition by using drift time information, which is the time between the positron crossing
and the avalanche formation on the anode wire: a basic configuration of three-cell ar-
rangement has been simulated through Monte Carlo techniques, then built and tested on
a prototype. Starting from three anode signals, a simple algorithm provides the track
parameters of a charged ionizing particle passing through the detector volume and the
single-hit resolution, defined as the uncertainty on the three impact parameters, is esti-
mated to be ∼ 100 μm.
In a stereo wires disposition, a drift cell mantains its shape but the transverse size changes,
in addition it is twisted along the z−axis direction: if the longitudinal cell extension is
∼ 2 m, the cell transverse size variation is non-negligible and gain variations effects arise.
Gain variations are studied in a 180 cm long single-cell prototype, built with the prelim-
inary set of wires: the experimental results are compared to Monte Carlo simulations.
Results obtained with these prototypes have been already implemented in the MEG II
drift chamber design: when the detector will be operative, they can be a valuable start-
ing point in achieving optimal features, both in terms of experimental resolutions and
functional stability
Deep neural network based audio source separation
No full textAudio source separation aims to extract individual sources from mixtures of multiple sound sources. Many techniques have been developed such as independent compo- nent analysis, computational auditory scene analysis, and non-negative matrix factorisa- tion. A method based on Deep Neural Networks (DNNs) and time-frequency (T-F) mask- ing has been recently developed for binaural audio source separation. In this method, the DNNs are used to predict the Direction Of Arrival (DOA) of the audio sources with respect to the listener which is then used to generate soft T-F masks for the recovery/estimation of the individual audio sources
Audio source separation with deep neural networks using the dropout algorithm
No full textA method based on Deep Neural Networks (DNNs) and time-frequency masking has been recently developed for binaural audio source separation. In this method, the DNNs are used to predict the Direction Of Arrival (DOA) of the audio sources with respect to the listener which is then used to generate soft time-frequency masks for the recovery/estimation of the individual audio sources. In this paper, an algorithm called ‘dropout’ will be applied to the hidden layers, affecting the sparsity of hidden units activations: randomly selected neurons and their connections are dropped during the training phase, preventing feature co-adaptation. These methods are evaluated on binaural mixtures generated with Binaural Room Impulse Responses (BRIRs), accounting a certain level of room reverberation. The results show that the proposed DNNs system with randomly deleted neurons is able to achieve higher SDRs performances compared to the baseline method without the dropout algorithm
Audio source separation with deep neural networks using the dropout algorithm
No full textA method based on Deep Neural Networks (DNNs) and time-frequency masking has been recently developed for binaural audio source separation. In this method, the DNNs are used to predict the Direction Of Arrival (DOA) of the audio sources with respect to the listener which is then used to generate soft time-frequency masks for the recovery/estimation of the individual audio sources. In this paper, an algorithm called ‘dropout’ will be applied to the hidden layers, affecting the sparsity of hidden units activations: randomly selected neurons and their connections are dropped during the training phase, preventing feature co-adaptation. These methods are evaluated on binaural mixtures generated with Binaural Room Impulse Responses (BRIRs), accounting a certain level of room reverberation. The results show that the proposed DNNs system with randomly deleted neurons is able to achieve higher SDRs performances compared to the baseline method without the dropout algorithm
Improving reverberant speech separation with binaural cues using temporal context and convolutional neural networks
No full textGiven binaural features as input, such as interaural level difference
and interaural phase difference, Deep Neural Networks (DNNs)
have been recently used to localize sound sources in a mixture of speech
signals and/or noise, and to create time-frequency masks for the estimation
of the sound sources in reverberant rooms. Here, we explore a
more advanced system, where feed-forward DNNs are replaced by Convolutional
Neural Networks (CNNs). In addition, the adjacent frames
of each time frame (occurring before and after this frame) are used to
exploit contextual information, thus improving the localization and separation
for each source. The quality of the separation results is evaluated
in terms of Signal to Distortion Ratio (SDR)
Binaural and log-power spectra features with deep neural networks for speech-noise separation
No full textBinaural features of interaural level difference and interaural phase difference have proved to be very effective in training deep neural networks (DNNs), to generate timefrequency masks for target speech extraction in speech-speech mixtures. However, effectiveness of binaural features is reduced in more common speech-noise scenarios, since the noise may over-shadow the speech in adverse conditions. In addition, the reverberation also decreases the sparsity of binaural features and therefore adds difficulties to the separation task. To address the above limitations, we highlight the spectral difference between speech and noise spectra and incorporate the log-power spectra features to extend the DNN input. Tested on two different reverberant rooms at different signal to noise ratios (SNR), our proposed method shows advantages over the baseline method using only binaural features in terms of signal to distortion ratio (SDR) and Short-Time Perceptual Intelligibility (STOI)
Development of an acoustic echo suppression model in an active audio system
No full textРоботу виконано на кафедрі комп'ютерних наук Тернопільського національного технічного університету імені Івана Пулюя. Захист відбудеться 27.06.2025р. на засіданні екзаменаційної комісії №30 у Тернопільському національному технічному університеті імені Івана ПулюяКваліфікаційна робота присвячена розробці алгоритму та моделі на його базі, котра здатна придушувати ехосигнали.
У першому розділі було поставлено та описано завдання на виконання дослідження, також були проведені оглядові роботи, пов'язані із тематикою дослідженням.
Другий розділ містить теоретичну частину, в якій вивчалися і аналізувалися алгоритми та методи, використані в роботі. Розроблено алгоритм на основі BLSTM, виходом якої є ідеальна бінарна маска. Ключовою особливістю запропонованого алгоритму є використання методів кластеризації (ЕМ, Mean-Shift, k-Means) на виході нейронної мережі.
У третьому розділі було описано та реалізовано запропоновану модель BLSTM+clustering. Проведено порівняння алгоритмів на сигналах бази даних TIMIT на основі загальноприйнятих метрик у обробці мовлення: ERLE, STOI, PESQ. Наведено результати експериментів та порівняння ефективності моделей. Показано, що використання кластеризації k-Means покращує роботу моделі BLSTM.
У четвертому розділі висвітлено важливі питання охорони праці та безпеки життєдіяльностіThe thesis deals with the development of an algorithm and a system based on it, which is capable of suppressing echo signals.
In the first chapter, the tasks for the research were set and described, and review works related to the research topic were also carried out.
The second chapter contains the theoretical part, in which the algorithms and methods used in the work were studied and analyzed. An algorithm based on BLSTM was developed, the output of which is an ideal binary mask. The key feature of the proposed algorithm is the use of clustering methods (EM, Mean-Shift, k-Means) at the output of the neural network.
In the third chapter, the proposed BLSTM+clustering model was described and implemented. A comparison of algorithms on signals from the TIMIT database was carried out based on generally accepted metrics in speech processing: ERLE, STOI, PESQ. The results of experiments and a comparison of the effectiveness of the models are presented. It is shown that the use of k-Means clustering improves the performance of the BLSTM model.
The fourth chapter highlights important issues of occupational health and safetyВСТУП 8 РОЗДІЛ 1. АНАЛІЗ ПРЕДМЕТНОЇ ОБЛАСТІ 10 1.1 Постановка завдання на проектування 10 1.2 Огляд аналогів 11 1.2.1 Огляд пов'язаних робіт 11 1.2.2 Відомі підходи 12 РОЗДІЛ 2. ТЕОРЕТИЧНА ЧАСТИНА 16 2.1 Аналіз акустичного еха 16 2.2 Рекурентна LSTM нейромережа 17 2.3 Алгоритм кластеризації К-Меаns 19 2.4 Короткочасне перетворення Фур'є 22 2.5 Метрики якості моделі 23 2.6 Вхідні дані для моделі 24 2.7 Вихід моделі 25 РОЗДІЛ 3. ПРАКТИЧНА ЧАСТИНА 26 3.1 Опис моделі BLSTM+clustering 26 3.2 Моделювання ехо-сигналу в приміщенні 27 3.2.1 Моделювання приміщення shoebox за допомогою методу ISM 28 3.2.2 Додавання джерел та мікрофонів 28 3.2.3 Створення імпульсної характеристики приміщення 29 3.2.4 Формування ехо-сигналу 30 3.3 Технології для реалізації моделі 31 3.4 Оцінка ефективності моделі BLSTM 32 3.4.1 Оцінка ефективності моделі BLSTM 32 3.4.2 Оцінка ефективності моделі BLSTM+k-Means 35 3.4.3 Оцінка ефективності моделі BLSTM+EM 37 3.4.4 Оцінка ефективності моделі BLSTM+Mean-Shift 38 3.4.5 Порівняння ефективності моделей 39 3.5 Оцінка ефективності моделей з однією розмовою 40 3.5.1 Оцінка ефективності моделі BLSTM 40 3.5.2 Оцінка ефективності моделі BLSTM +К-Means 42 3.5.3 Порівняння ефективності моделей 44 РОЗДІЛ 4. БЕЗПЕКА ЖИТТЄДІЯЛЬНОСТІ, ОСНОВИ ОХОРОНИ ПРАЦІ 45 4.1 Класифікація шкідливих та небезпечних виробничих факторів 45 4.2 Вплив вібрації на людину 49 ВИСНОВКИ 53 ПЕРЕЛІК ДЖЕРЕЛ 54 ДОДАТК
