Model size and complexity remain the biggest challenges in the deployment of speech enhancement and separation systems on low-resource devices such as earphones and hearing aids. Although methods such as compression, distillation and quantization can be applied to large models, they often come with a cost on the model performance. In this paper, we provide a simple model design paradigm that explicitly designs ultra-lightweight models without sacrificing the performance. Motivated by the sub-band frequency-LSTM (F-LSTM) architectures, we introduce the group communication (GroupComm), where a feature vector is split into smaller groups and a small processing block is used to perform inter-group communication. Unlike standard F-LSTM models where the sub-band outputs are concatenated, an ultra-small module is applied on all the groups in parallel, which allows a significant decrease on the model size. Experiment results show that comparing with a strong baseline model which is already lightweight, GroupComm can achieve on par performance with 35.6 times fewer parameters and 2.3 times fewer operations.
In this paper we propose to bridge the gap between using extremely low resolution 1-bit measurements and estimating targets' parameters, such as their velocities, that exist in a continuum, i.e., by performing Off-the-Grid estimation. To that end, a Continuous version of Orthogonal Matching Pursuit (COMP) is modified in order to leverage the 1-bit measurements coming from a simple Doppler Radar. Using Monte-Carlo simulations, we show that although the resolution of the acquisition is dramatically reduced, velocity estimation of multiple targets can still be achieved and reaches performances beyond classic On-the-Grid methods. Furthermore, we show empirically that adding a random and uniform dithering before the quantization is necessary when estimating more than one target.
Reset control is known to be able to outperform PID and the like linear controllers. However, in motion control systems, quantization can cause severe performance degradation. This paper shows the application of time regularization to mitigate this practical issue in reset control systems. Numerical simulations have been conducted in order to analyze the cause of the quantization induced performance degradation and the effectiveness of time regularization to mitigate this degradation; with tuning guidelines for the time regularization parameter also provided. Moreover, a robustness analysis is performed. The solution is also tested experimentally on a high precision motion system for validation. It is estimated by numerical simulations that time regularization can reduce quantization induced performance degradation by almost 10 dB. Experiments have similarly shown a reduction of several dB for the high precision motion stage.
Cell-free (CF) massive multiple-input-multiple-output (mMIMO) deployments are usually investigated with half-duplex nodes and high-capacity fronthaul links. To leverage the possible gains in throughput and energy efficiency (EE) of full-duplex (FD) communications, we consider a FD CF mMIMO system with practical limited-capacity fronthaul links. We derive closed-form spectral efficiency (SE) lower bounds for this system with maximum-ratio combining/maximum-ratio transmission processing and optimal uniform quantization. We then optimize the weighted sum EE (WSEE) via downlink and uplink power control by using a {two-layered} approach: the first layer formulates the optimization as a generalized convex program, while the second layer solves the optimization decentrally using alternating direction method of multipliers. We analytically show that the proposed two-layered formulation yields a Karush-Kuhn-Tucker point of the original WSEE optimization. We numerically show the influence of weights on the individual EE of the users, which demonstrates the utility of WSEE metric to incorporate heterogeneous EE requirements of users. We show that the low fronthaul capacity reduces the number of users each AP can support, and the cell-free system, consequently, becomes user-centric.
One of the main drawbacks of the well-known Direct Position Determination (DPD) method is the requirement that raw signal data be transferred to a common processor. It would therefore be of high practical value if DPD$-$or a modified version thereof$-$could be successfully applied to a coarsely quantized version of the raw data, thus alleviating the requirements on the communication links between the different base stations. Motivated by the above, and inspired by recent work in the rejuvenated one-bit array processing field, we present One-Bit DPD: a method for direct localization based on one-bit quantized measurements. We show that despite the coarse quantization, the proposed method nonetheless yields an estimate for the unknown emitter position with appealing asymptotic properties. We further establish the underlying identifiability conditions of this model, which rely only on second-order statistics. Empirical simulation results corroborate our analytical derivations, demonstrating that much of the information regarding the unknown emitter position is preserved under this crude form of quantization.
One-bit quantization has recently become an attractive option for data acquisition in cutting edge applications, due to the increasing demand for low power and higher sampling rates. Subsequently, the rejuvenated one-bit array processing field is now receiving more attention, as "classical" array processing techniques are adapted / modified accordingly. However, array calibration, often an instrumental preliminary stage in array processing, has so far received little attention in its one-bit form. In this paper, we present a novel solution approach for the blind calibration problem, namely, without using known calibration signals. In order to extract information within the second-order statistics of the quantized measurements, we propose to estimate the unknown sensors' gains and phases offsets according to a Kullback-Leibler Divergence (KLD) covariance fitting criterion. We then provide a quasi-Newton solution algorithm, with a consistent initial estimate, and demonstrate the improved accuracy of our KLD-based estimates in simulations.
Cell-free (CF) massive multiple-input-multiple-output (mMIMO) deployments are usually investigated with half-duplex nodes and high-capacity fronthaul links. To leverage the possible gains in throughput and energy efficiency (EE) of full-duplex (FD) communications, we consider a FD CF mMIMO system with practical limited-capacity fronthaul links. We derive closed-form spectral efficiency (SE) lower bounds for this system with maximum-ratio combining/maximum-ratio transmission processing and optimal uniform quantization. We then optimize the weighted sum EE (WSEE) via downlink and uplink power control by using a two-layered approach: the first layer formulates the optimization as a generalized convex program, while the {second layer} solves the optimization decentrally using the alternating direction method of multipliers. We analytically show that the proposed two-layered formulation yields a Karush-Kuhn-Tucker point of the original WSEE optimization. We numerically show the influence of weights on the individual EE of the users, which demonstrates the utility of the WSEE metric to incorporate heterogeneous EE requirements of users. We show that low fronthaul capacity reduces the number of users each AP can support, and the cell-free system, consequently, becomes user-centric.
This paper focuses on channel estimation for mmWave MIMO systems with 1-bit spatial sigma-delta analog-to-digital converters (ADCs). The channel estimation performance with 1-bit spatial sigma-delta ADCs depends on the quantization noise modeling. Therefore, we present a new method for modeling the quantization noise by leveraging the deterministic input-output relation of the 1-bit spatial sigma-delta ADC. Using this new noise model, we propose an algorithm for channel estimation for a narrowband single-user mmWave line-of-sight MIMO system by determining the unknown angles and path attenuation that characterize the flat fading channel. Through simulations, we demonstrate that the performance of the developed method is comparable to the traditional analog systems and significantly better than the conventional 1-bit quantized systems.
Measuring the quality of digital videos viewed by human observers has become a common practice in numerous multimedia applications, such as adaptive video streaming, quality monitoring, and other digital TV applications. Here we explore a significant, yet relatively unexplored problem: measuring perceptual quality on videos arising from both luma and chroma distortions from compression. Toward investigating this problem, it is important to understand the kinds of chroma distortions that arise, how they relate to luma compression distortions, and how they can affect perceived quality. We designed and carried out a subjective experiment to measure subjective video quality on both luma and chroma distortions, introduced both in isolation as well as together. Specifically, the new subjective dataset comprises a total of $210$ videos afflicted by distortions caused by varying levels of luma quantization commingled with different amounts of chroma quantization. The subjective scores were evaluated by $34$ subjects in a controlled environmental setting. Using the newly collected subjective data, we were able to demonstrate important shortcomings of existing video quality models, especially in regards to chroma distortions. Further, we designed an objective video quality model which builds on existing video quality algorithms, by considering the fidelity of chroma channels in a principled way. We also found that this quality analysis implies that there is room for reducing bitrate consumption in modern video codecs by creatively increasing the compression factor on chroma channels. We believe that this work will both encourage further research in this direction, as well as advance progress on the ultimate goal of jointly optimizing luma and chroma compression in modern video encoders.
Quantization of signals is an integral part of modern signal processing applications, such as sensing, communication, and inference. While signal quantization provides many physical advantages, it usually degrades the subsequent estimation performance that is based on quantized data. In order to maintain physical constraints and simultaneously bring substantial performance gain, in this work we consider systems with mixed-resolution, 1-bit quantized and continuous-valued, data. First, we describe the linear minimum mean-squared error (LMMSE) estimator and its associated mean-squared error (MSE) for the general mixed-resolution model. However, the MSE of the LMMSE requires matrix inversion in which the number of measurements defines the matrix dimensions and thus, is not a tractable tool for optimization and system design. Therefore, we present the linear Gaussian orthonormal (LGO) measurement model and derive a closed-form analytic expression for the MSE of the LMMSE estimator under this model. In addition, we present two common special cases of the LGO model: 1) scalar parameter estimation and 2) channel estimation in mixed-ADC multiple-input multiple-output (MIMO) communication systems. We then solve the resource allocation optimization problem of the LGO model with the proposed tractable form of the MSE as an objective function and under a power constraint using a one-dimensional search. Moreover, we present the concept of dithering for mixed-resolution models and optimize the dithering noise as part of the resource allocation optimization problem for two dithering schemes: 1) adding noise only to the quantized measurements and 2) adding noise to both measurement types. Finally, we present simulations that demonstrate the advantages of using mixed-resolution measurements and the possible improvement introduced with dithering and resource allocation.
This paper proposes an adaptive random experiment design (ARED) algorithm that can be applied to optimize the multiple factors and levels experiments. The algorithm takes real-time model error as the adaptive condition, and outputs a model that conforms to the error quantization standard based on the automatic process. According to the actual experimental scenario, the similar number of test cases were selected between the ARED method and the comparative experimental design method under the bimodal Gaussian function, the bimodal surface function and the peaks function, respectively. simultaneously, the support vector machine (SVM) algorithm is used to construct the model for the selected test cases, and the verification surface (or curve) is predicted. The qualitative and quantitative analysis is carried out at two-slice of applicability and precision. The results show that the ARED method can be applied to the experiment of multi-factor, and has better precision and applicability than the comparative experimental methods.
Recently, sequence-to-sequence models with attention have been successfully applied in Text-to-speech (TTS). These models can generate near-human speech with a large accurately-transcribed speech corpus. However, preparing such a large data-set is both expensive and laborious. To alleviate the problem of heavy data demand, we propose a novel unsupervised pre-training mechanism in this paper. Specifically, we first use Vector-quantization Variational-Autoencoder (VQ-VAE) to ex-tract the unsupervised linguistic units from large-scale, publicly found, and untranscribed speech. We then pre-train the sequence-to-sequence TTS model by using the<unsupervised linguistic units, audio>pairs. Finally, we fine-tune the model with a small amount of<text, audio>paired data from the target speaker. As a result, both objective and subjective evaluations show that our proposed method can synthesize more intelligible and natural speech with the same amount of paired training data. Besides, we extend our proposed method to the hypothesized low-resource languages and verify the effectiveness of the method using objective evaluation.
We propose a deep learning-based channel estimation, quantization, feedback, and precoding method for downlink multiuser multiple-input and multiple-output systems. In the proposed system, channel estimation and quantization for limited feedback are handled by a receiver deep neural network (DNN). Precoder selection is handled by a transmitter DNN. To emulate the traditional channel quantization, a binarization layer is adopted at each receiver DNN, and the binarization layer is also used to enable end-to-end learning. However, this can lead to inaccurate gradients, which can trap the receiver DNNs at a poor local minimum during training. To address this, we consider knowledge distillation, in which the existing DNNs are jointly trained with an auxiliary transmitter DNN. The use of an auxiliary DNN as a teacher network allows the receiver DNNs to additionally exploit lossless gradients, which is useful in avoiding a poor local minimum. For the same number of feedback bits, our DNN-based precoding scheme can achieve a higher downlink rate compared to conventional linear precoding with codebook-based limited feedback.
For dynamic visible light communication (VLC) systems, received optical power fluctuates largely due to the fast movement of VLC terminals. Such fluctuations will cause possible signal clipping and quantization noise at the analog-to-digital converters (ADCs) and thus aggravate signal-to-noise ratio (SNR) and reliability of the communication link. To mitigate this effect, the automatic gain control (AGC) technique is often introduced in the receiver front-end to adaptively adjust the electrical signal strength. In this paper, we provided a VLC front-end analysis considering AGC model. Based on the model, some design principles of AGC are theoretically derivated: the effects of AGC index $m$, AGC maximum gain $g_{max}$ and AGC equilibrium range $DR$. Next, an analog AGC amplifier was carefully implemented and the effectiveness of our AGC modeling was proved through BER experiments. Finally, realtime 25Mb/s on-off-keying (OOK) experiments with AGC function were demonstrated in a dynamic link with different speeds. Experimental results show that the AGC can stabilize the BER performance and improve the system performance in the speed of 1m/s.
We propose an original method for vectorizing an image or zooming it at an arbitrary scale. The core of our method relies on the resolution of a geometric variational model and therefore offers theoretic guarantees. More precisely, it associates a total variation energy to every valid triangulation of the image pixels. Its minimization induces a trian-gulation that reflects image gradients. We then exploit this triangulation to precisely locate discontinuities, which can then simply be vectorized or zoomed. This new approach works on arbitrary images without any learning phase. It is particularly appealing for processing images with low quantization like pixel art and can be used for depixelizing such images. The method can be evaluated with an online demonstrator, where users can reproduce results presented here or upload their own images.
Time Divison Duplex (TDD) wireless communication systems are inherently bidirectional, which facilitates exploiting channel reciprocity for pilot based channel estimation of both uplink and downlink. However, there exists a gross asymmetry in channel estimation complexity for the uplink and downlink, particularly for Multiple Input Multiple Output (MIMO) TDD systems. Usually, Base Stations (BS) with more antennas need to estimate fewer parameters from each antenna, whereas the estimation requirement is disproportionately higher at the User Equipment (UE). Unlike the UE, the BS has powerful hardware, computational resources and energy to accurately estimate and track channel profiles. To overcome this asymmetry, we propose a solution for MIMO-TDD downlink communication, wherein the post-coder part of the channel matrix is quantized at the BS, and is communicated to the UE via a low-rate channel. Using asymptotically tight lower bounds on the downlink achievable rates, we quantify the performance of the proposed scheme. Simulations reveal that a moderate number of quantization bits are sufficient to achieve rates close to the the link capacity. We further show that, when the BS has many more antennas at than the UE, the channel can be compensated by appropriate transmit domain precoding without post-coder knowledge at the UE.
Analog-to-digital converters (ADCs) allow physical signals to be processed using digital hardware. The power consumed in conversion grows with the sampling rate and quantization resolution, imposing a major challenge in power-limited systems. A common ADC architecture is based on sample-and-hold (S/H) circuits, where the analog signal is being tracked only for a fraction of the sampling period. In this paper, we propose the concept of eSampling ADCs, which harvest energy from the analog signal during the time periods where the signal is not being tracked. This harvested energy can be used to supplement the ADC itself, paving the way to the possibility of zero-power consumption and power-saving ADCs. We analyze the tradeoff between the ability to recover the sampled signal and the energy harvested, and provide guidelines for setting the sampling rate in the light of accuracy and energy constraints. Our analysis indicates that eSampling ADCs operating with up to 12 bits per sample can acquire bandlimited analog signals such that they can be perfectly recovered without requiring power from the external source. Furthermore, our theoretical results reveal that eSampling ADCs can in fact save power by harvesting more energy than they consume. To verify the feasibility of eSampling ADCs, we present a circuit-level design using standard complementary metal oxide semiconductor (CMOS) 65 nm technology. An eSampling 8-bit ADC which samples at 40 MHZ is designed on a Cadence Virtuoso platform. Our experimental study involving Nyquist rate sampling of bandlimited signals demonstrates that such ADCs are indeed capable of harvesting more energy than that spent during analog-to-digital conversion, without affecting the accuracy.
Recursive marginal quantization (RMQ) allows the construction of optimal discrete grids for approximating solutions to stochastic differential equations in d-dimensions. Product Markovian quantization (PMQ) reduces this problem to d one-dimensional quantization problems by recursively constructing product quantizers, as opposed to a truly optimal quantizer. However, the standard Newton-Raphson method used in the PMQ algorithm suffers from numerical instabilities, inhibiting widespread adoption, especially for use in calibration. By directly specifying the random variable to be quantized at each time step, we show that PMQ, and RMQ in one dimension, can be expressed as standard vector quantization. This reformulation allows the application of the accelerated Lloyd's algorithm in an adaptive and robust procedure. Furthermore, in the case of stochastic volatility models, we extend the PMQ algorithm by using higher-order updates for the volatility or variance process. We illustrate the technique for European options, using the Heston model, and more exotic products, using the SABR model.
In this paper, by adopting the superposition of terms with unequal coefficients, some useful closed-form formulas for the class of large planar arrays are presented for predicting the directivities of radiated beams. Despite the applied simplifying assumptions, the provided formulas may be used for designing a large multiple-beam planar array with a specified power pattern. Several illustrative examples are presented which are numerically demonstrate by the MATLAB software. The applicability of the proposed synthesis method verified by the achieved conformity between the simulations and theoretical predictions. Furthermore, the impact of quantization and array dimensions on the performance of power controlling as well as the limitations on the maximum scan angle are investigated. It is believed that the proposed straightforward approach for array synthesis provides a new opportunity for various applications such as multiple-target radar systems and beamforming array antenna.
Robust automatic speech recognition (ASR) system exploits state-of-the-art deep neural networks (DNN) based acoustic model (AM) trained with Lattice Free-Maximum Mutual Information (LF-MMI) criterion and n-gram language models. These systems are quite large and require significant parameter reduction to operate on embedded devices. Impact of the parameter quantization on the overall word recognition performance is studied in this paper. Following three approaches are presented: (i) AM trained in Kaldi framework with conventional factorized TDNN (TDNN-F) architecture. (ii) the TDNN built in Kaldi is loaded into the Pytorch toolkit using a C++ wrapper. The weights and activation parameters are then quantized and the inference is performed in Pytorch. (iii) post quantization training for fine-tuning. Results obtained on standard Librispeech setup provide an interesting overview of recognition accuracy w.r.t. applied quantization scheme.