We present a method to maintain the subjective perception of volume of audio signals and, at the same time, reduce their absolute peak value. We focus on achieving this without compromising the perceived audio quality. This is specially useful, for example, to maximize the perceived reproduction level of loudspeakers where simply amplifying the signal amplitude, and hence their peak value, is limited due to already constrained physical designs. In particular, we minimize the absolute peak value subject to a constraint based on auditory masking. This limits the perceptual difference between the original and the modified signals. Moreover, this constraint can be tuned and allows to control the resulting audio quality. We show results comparing loudness and audio quality as a function of peak reduction. These results suggest that our method presents the best trade-off between loudness and audio quality when compared against classical methods based on compression and clipping.

In communication systems, it is crucial to estimate the perceived quality of audio and speech. The industrial standards for many years have been PESQ, 3QUEST, and POLQA, which are intrusive methods. This restricts the possibilities of using these metrics in real-world conditions, where we might not have access to the clean reference signal. In this work, we develop a new non-intrusive metric based on crowd-sourced data. We build a new speech dataset by combining publicly available speech, noises, and reverberations. Then we follow the ITU P.808 recommendation to label the dataset with mean opinion scores (MOS). Finally, we train a deep neural network to estimate the MOS from the speech data in a non-intrusive way. We propose two novelties in our work. First, we explore transfer learning by pre-training a model using a larger set of POLQA scores and finetuning with the smaller (and thus cheaper) human-labeled set. Secondly, we perform a subject-specific scaling in the MOS scores to adjust for their different subjective scales. Our model yields better accuracy than PESQ, POLQA, and other non-intrusive methods when evaluated on the independent VCTK test set. We also report misleading POLQA scores for reverberant speech.

In this paper we describe our approach to teaching applied digital signal processing (DSP) using freely-available software and inexpensive off-the-shelf hardware components. The pedagogical framework is built around simple real-time audio processing algorithms that provide immediate and engaging feedback to the students. At the same time, our end goal is to build a learning module that can be easily reproduced by fellow instructors and used in our Massive Open Online Course (MOOC). We use Python for the initial algorithmic prototyping and then transition to a hardware implementation using an STMicroelectronics' Nucleo-64 core with Adafruit I/O breakout boards. We employ modern documentation tools based on Git and Markdown in order to share the material and the exercises, allowing students and educators to access, reproduce, and contribute changes to the content.

Stained glass windows are designed to reveal their powerful artistry under diverse and time-varying lighting conditions; virtual relighting of stained glass, therefore, represents an exceptional tool for the appreciation of this age old art form. However, as opposed to most other artifacts, stained glass windows are extremely difficult if not impossible to analyze using controlled illumination because of their size and position. In this paper, we present novel methods built upon image based priors to perform virtual relighting of stained glass artwork by acquiring the actual light transport properties of a given artifact. In a preprocessing step, we build a material-dependent dictionary for light transport by studying the scattering properties of glass samples in a laboratory setup. We can now use the dictionary to recover a light transport matrix in two ways: under controlled illuminations the dictionary constitutes a sparsifying basis for a compressive sensing acquisition, while in the case of uncontrolled illuminations the dictionary is used to perform sparse regularization. The proposed basis preserves volume impurities and we show that the retrieved light transport matrix is heterogeneous, as in the case of real world objects. We present the rendering results of several stained glass artifacts, including the Rose Window of the Cathedral of Lausanne, digitized using the presented methods.

After years of experimentation, online teaching has gone through a phase transition with the appearance of massive open online courses (MOOCs), following the model pioneered by Salman Khan with short videos and the use of tablets. Dozens of university-level courses are now available, and followed by hundreds of thousands of online students. We report on early experiences with teaching “Fundamentals of Electrical Engineering” and “Digital Signal Processing” on an open online platform (Coursera). We address in particular: (i) suitability of online platforms for signal processing oriented topics; (ii) structuring of material for the online format; (iii) quizzes and exercises for large classes; (iv) grading methods and possible certification issues; (v) learning from teaching, forums, and data mining of students feedback.

Reducing energy cost is crucial for energy-constrained smart wireless cameras. Existing platforms impose two main challenges: First, most commercial smart phones have a closed platform, which makes it impossible to manage low-level circuits. Since the sampling frequency is moderate in environmental monitoring context, any improper power management in idle period will incur significant energy leak. Secondly, low-end cameras tailored for wireless sensor networks usually have limited processing power or communication range, and thus are not capable of outdoor monitoring task under low data rate. To tackle these issues, we develop Sensorcam, a long-range, smart wireless camera running a Linux-base open system. Through better power management in idle period and the "intelligence" of the camera itself, we demonstrate an energy-efficient wireless monitoring system in a real deployment.

A common technique to extend linear prediction to nonstationary signals is time segmentation: the signal is split into small portions and the modelization is carried out locally. The accuracy of the analysis is, however, dependent on the window size and on the signal characteristics, so that the problem of finding a good segmentation is crucial to the entire modeling scheme. In this paper, we present an algorithm which determines the optimal segmentation with respect to a cost function relating prediction error to modeling cost. The proposed approach casts the problem in a rate/distortion (R/D) framework, whereby the segmentation is implicitly computed while minimizing the modelization distortion for a given modelization cost. The algorithm is implemented by means of dynamic programming and takes the form of a trellis-based Lagrangian minimization. The optimal linear predictor, when applied to speech coding, dramatically reduces the number of bits per second devoted to the modeling parameters in comparison to fixed-window schemes.

For a given set of quantizers and a data vector, the optimal bit allocation in a rate/distortion sense is the sequence of quantizers which minimizes the overall distortion for a given bit budget. In an operational framework, this sequence is dependent on the data realization rather than on its probabilistic model and the cost of describing the sequence itself becomes therefore part of the bit budget. We present an allocation algorithm based on dynamic programming which determines the optimal bit allocation taking into account the side information of describing the structure of the allocation itself; practical simplifications of the algorithm are also presented with respect to coding of continuous data sources.

We argue that Java is a natural language to develop interactive teaching material that can be shared and distributed widely. Unlike any other programming language or platform we know, Java development is justified because of its almost universal acceptance. We develop a block diagram (BD) based approach that allows one to develop interactive and downloadable signal processing laboratories. As an example, we show how specific experiments for a DSP class, as well as for an advanced course on wavelets have been developed. The article first explains why the Java language has been chosen, and then describes what has been realized today. Finally, we show how the BD representation can be efficiently used for the development of a wavelet theory course. It is shown that only a few simple blocks are sufficient for creating many didactic programs. This can be seen as an a posteriori justification of the BD model.

A data transmission framework is proposed to embed digital data into an audio signal in a perceptually undetectable or almost undetectable way. The resulting signal can be reproduced as is with no loss of acoustic quality; the embedded data can be exactly retrieved at the decoder. The transmission process exploits the perceptual redundancy of the audio signal to conceal the acoustic impact of the embedded data; encoding of side information is used to inform the receiver of the time-varying structure of the masking properties of the audio signal. A sample implementation is described with a throughput of the order of 30 kbit/sec over CD-quality audio.

The idea of optimal joint time segmentation and resource allocation for signal modeling is explored with respect to arbitrary segmentations and arbitrary representation schemes. When the chosen signal modeling techniques can be quantified in terms of a cost function which is additive over distinct segments, a dynamic programming approach guarantees the global optimality of the scheme while keeping the computational requirements of the algorithm sufficiently low. Two immediate applications of the algorithm to LPC speech coding and to sinusoidal modeling of musical signals are presented.