3,027 research outputs found
Synthetic Training Set Generation using Text-To-Audio Models for Environmental Sound Classification
No full textIn recent years, text-to-audio models have revolutionized the field of automatic audio generation. This paper investigates their application in generating synthetic datasets for training data-driven models. Specifically, this study analyzes the performance of two environmental sound classification systems trained with data generated from text-to-audio models. We considered three scenarios: a) augmenting the training dataset with data generated by text-to-audio models; b) using a mixed training dataset combining real and synthetic text-driven generated data; and c) using a training dataset composed entirely of synthetic audio. In all cases, the performance of the classification models was tested on real data. Results indicate that text-to-audio models are effective for dataset augmentation, with consistent performance when replacing a subset of the recorded dataset. However, the performance of the audio recognition models drops when relying entirely on generated audio
An assessment of the impact of possible CAP reform scenarios on Romanian agriculture
No full textUsing a simplified model, with key-variable the prices of two different possible scenarios of CAP reform after 2013 (moderate and radical), this paper present a comparison between the price effects of implementation of each reform scenario at 2015 horizon on Romanian agriculture. This short analysis shows that, under the presented hypotheses, the net welfare effect, due to the price changes, for the selected products, is positive in both reform scenarios, yet greater in the case of the radical reform. Integrated in the large context of Romanian development, it seems that the influence of CAP reform upon agriculture and rural areas will be most likely a gradual one: an interpenetration between the two scenarios is foreseeable, starting with the moderate reform that will dominate the period around 2013, the reform measures acquiring a more radical character afterwards.CAP reform, Romania, welfare effects, Agricultural and Food Policy,
Rich, Sturmian, and trapezoidal words
No full textIn this paper we explore various interconnections between rich words, Sturmian words, and trapezoidal words. Rich words, first introduced by the second and third authors together with J. Justin and S. Widmer, constitute a new class of finite and infinite words characterized by having the maximal number of palindromic factors. Every finite Sturmian word is rich, but not conversely. Trapezoidal words were first introduced by the first author in studying the behavior of the subword complexity of finite Sturmian words. Unfortunately this property does not characterize finite Sturmian words. In this note we show that the only trapezoidal palindromes are Sturmian. More generally we show that Sturmian palindromes can be characterized either in terms of their subword complexity (the trapezoidal property) or in terms of their palindromic complexity. We also obtain a similar characterization of rich palindromes in terms of a relation between palindromic complexity and subword complexity
Deep Prior-Based Audio Inpainting Using Multi-Resolution Harmonic Convolutional Neural Networks
No full textIn this manuscript, we propose a novel method to perform audio inpainting, i.e., the restoration of audio signals presenting multiple missing parts. Audio inpainting can be interpreted in the context of inverse problems as the task of reconstructing an audio signal from its corrupted observation. For this reason, our method is based on a deep prior approach, a recently proposed technique that proved to be effective in the solution of many inverse problems, among which image inpainting. Deep prior allows one to consider the structure of a neural network as an implicit prior and to adopt it as a regularizer. Differently from the classical deep learning paradigm, deep prior performs a single-element training and thus it can be applied to corrupted audio signals independently from the available training data sets. In the context of audio inpainting, a network presenting relevant audio priors will possibly generate a restored version of an audio signal, only provided with its corrupted observation. Our method exploits a time-frequency representation of audio signals and makes use of a multi-resolution convolutional autoencoder, that has been enhanced to perform the harmonic convolution operation. Results show that the proposed technique is able to provide a coherent and meaningful reconstruction of the corrupted audio. It is also able to outperform the methods considered for comparison, in its domain of application
Characterization Results for the Poset Based Representation of Topological Relations - I: Introduction and Models
No full text@article{DBLP:journals/informaticaSI/ForlizziN99,
author = {Luca Forlizzi and
Enrico Nardelli},
title = {Characterization Results for the Poset Based Representation
of Topological Relations - I: Introduction and Models.},
journal = {Informatica (Slovenia)},
volume = {23},
number = {2},
year = {1999},
bibsource = {DBLP, http://dblp.uni-trier.de}
Characterization Results for the Poset Based Representation of Topological Relations - II: Intersection and Union
No full text@article{DBLP:journals/informaticaSI/ForlizziN00,
author = {Luca Forlizzi and
Enrico Nardelli},
title = {Characterization Results for the Poset Based Representation
of Topological Relations - II: Intersection and Union.},
journal = {Informatica (Slovenia)},
volume = {24},
number = {1},
year = {2000},
bibsource = {DBLP, http://dblp.uni-trier.de}
System-on-chip Computing and Interconnection Architectures for Telecommunications and Signal Processing
No full textThis dissertation proposes novel architectures and design techniques targeting SoC building blocks for telecommunications and signal processing applications.
Hardware implementation of Low-Density Parity-Check decoders is approached at both the algorithmic and the architecture level. Low-Density Parity-Check codes are a promising coding scheme for future communication standards due to their outstanding error correction performance.
This work proposes a methodology for analyzing effects of finite precision arithmetic on error correction performance and hardware complexity. The methodology is throughout employed for co-designing the decoder. First, a low-complexity check node based on the P-output decoding principle is designed and characterized on a CMOS standard-cells library. Results demonstrate implementation loss below 0.2 dB down to BER of 10^{-8} and a saving in complexity up to 59% with respect to other works in recent literature. High-throughput and low-latency issues are addressed with modified single-phase decoding schedules. A new "memory-aware" schedule is proposed requiring down to 20% of memory with respect to the traditional two-phase flooding decoding. Additionally, throughput is doubled and logic complexity reduced of 12%. These advantages are traded-off with error correction performance, thus making the solution attractive only for long codes, as those adopted in the DVB-S2 standard. The "layered decoding" principle is extended to those codes not specifically conceived for this technique. Proposed architectures exhibit complexity savings in the order of 40% for both area and power consumption figures, while implementation loss is smaller than 0.05 dB.
Most modern communication standards employ Orthogonal Frequency Division Multiplexing as part of their physical layer. The core of OFDM is the Fast Fourier Transform and its inverse in charge of symbols (de)modulation. Requirements on throughput and energy efficiency call for FFT hardware implementation, while ubiquity of FFT suggests the design of parametric, re-configurable and re-usable IP hardware macrocells. In this context, this thesis describes an FFT/IFFT core compiler particularly suited for implementation of OFDM communication systems. The tool employs an accuracy-driven configuration engine which automatically profiles the internal arithmetic and generates a core with minimum operands bit-width and thus minimum circuit complexity. The engine performs a closed-loop optimization over three different internal arithmetic models (fixed-point, block floating-point and convergent block floating-point) using the numerical accuracy budget given by the user as a reference point. The flexibility and re-usability of the proposed macrocell are illustrated through several case studies which encompass all current state-of-the-art OFDM communications standards (WLAN, WMAN, xDSL, DVB-T/H, DAB and UWB). Implementations results are presented for two deep sub-micron standard-cells libraries (65 and 90 nm) and commercially available FPGA devices. Compared with other FFT core compilers, the proposed environment produces macrocells with lower circuit complexity and same system level performance (throughput, transform size and numerical accuracy).
The final part of this dissertation focuses on the Network-on-Chip design paradigm whose goal is building scalable communication infrastructures connecting hundreds of core. A low-complexity link architecture for mesochronous on-chip communication is discussed. The link enables skew constraint looseness in the clock tree synthesis, frequency speed-up, power consumption reduction and faster back-end turnarounds. The proposed architecture reaches a maximum clock frequency of 1 GHz on 65 nm low-leakage CMOS standard-cells library. In a complex test case with a full-blown NoC infrastructure, the link overhead is only 3% of chip area and 0.5% of leakage power consumption.
Finally, a new methodology, named metacoding, is proposed. Metacoding generates correct-by-construction technology independent RTL codebases for NoC building blocks. The RTL coding phase is abstracted and modeled with an Object Oriented framework, integrated within a commercial tool for IP packaging (Synopsys CoreTools suite). Compared with traditional coding styles based on pre-processor directives, metacoding produces 65% smaller codebases and reduces the configurations to verify up to three orders of magnitude
Towards HRTF Personalization using Denoising Diffusion Models
No full textHead-Related Transfer Functions (HRTFs) have fundamental applications for realistic rendering in immersive audio scenarios. However, they are strongly subject-dependent as they vary considerably depending on the shape of the ears, head and torso. Thus, personalization procedures are required for accurate binaural rendering. Recently, Denoising Diffusion Probabilistic Models (DDPMs), a class of generative learning techniques, have been applied to solve a variety of signal processing-related problems. In this paper, we propose a first approach for using DDPM conditioned on anthropometric measurements to generate personalized Head-Related Impulse Response (HRIR), the time-domain representation of HRTF. The results show the feasibility of DDPMs for HRTF personalization obtaining performance in line with state-of-the-art models
A Fast Ray Space Transform for Wave Field Processing using Acoustic Arrays
No full textThe importance of soundfield imaging techniques is expected to further increase in the next few years thanks to the ever-increasing availability of low-cost sensors such as MEMS microphones. When it comes to processing a relevant number of sensor signals, however, the computational load of space-time processing algorithms easily grows to unmanageable levels. The Ray Space Transform (RST) was recently introduced as a promising tool for soundfield analysis. Given the collection of signals captured by a uniform linear array of microphones, the RST allows us to collect and map the directional components of the acoustic field onto a domain called “ray space”, where relevant acoustic objects are represented as linear patterns for advanced acoustic analysis and synthesis applications. So far the computational complexity of the RST linearly increases with the number of microphones. In order to alleviate this problem, in this paper we propose an alternative efficient implementation of the RST based on the Non Uniform Fast Fourier Transform
REAL-TIME MULTICHANNEL SPEECH SEPARATION AND ENHANCEMENT USING A BEAMSPACE-DOMAIN-BASED LIGHTWEIGHT CNN
No full textThe problems of speech separation and enhancement concern the extraction of the speech emitted by a target speaker when placed in a scenario where multiple interfering speakers or noise are present, respectively. A plethora of practical applications such as home assistants and teleconferencing require some sort of speech separation and enhancement pre-processing before applying Automatic Speech Recognition (ASR) systems. In the recent years, most techniques have focused on the application of deep learning to either time-frequency or time-domain representations of the input audio signals. In this paper we propose a real-time multichannel speech separation and enhancement technique, which is based on the combination of a directional representation of the sound field, denoted as beamspace, with a lightweight Convolutional Neural Network (CNN). We consider the case where the Direction-Of-Arrival (DOA) of the target speaker is approximately known, a scenario where the power of the beamspace-based representation can be fully exploited, while we make no assumption regarding the identity of the talker. We present experiments where the model is trained on simulated data and tested on real recordings and we compare the proposed method with a similar state-of-the-art technique
- …
