Majid Fozunbal

On the quality-assessment of reverberated speech

This paper addresses the problem of quantifying the reverberation effect in speech signals. The perception of reverberation is assessed based on a new measure combining the characteristics of reverberation time, room spectral variance, and direct-to-reverberant energy ratio, which are estimated from the associated room impulse response (RIR). The practical aspects behind a robust RIR estimation are underlined, allowing an effective feature extraction for reverberation evaluation. The resulting objective metric achieves a correlation factor of about 90% with the subjective scores of two distinct speech databases, illustrating the systems ability to assess the reverberation effect in a reliable manner.

Feature analysis for quality assessment of reverberated speech

This paper analyzes the ability of several measurements to quantify the reverberation effect in speech signals. We consider an intrusive scheme, in which the clean and reverberated signals are available, allowing one to estimate the corresponding room impulse response (RIR) signal. An artificial neural network (ANN) is trained for all features and used in a regression approach to estimate the human perceptual evaluation in a mean opinion score (MOS) 1–5 scale. Dimensionality reduction approaches are applied to generate a simpler ANN regression, establishing the most representative features for the problem at hand. A correlation level of 85% with subjective test scores was achieved by reducing the input-vector dimension from 10 to 3, including only the features of reverberation time, room spectral variance, and direct-to-reverberant energy ratio.

On regret of parametric mismatch in minimum mean square error estimation

This paper studies the effect of parametric mis-match in minimum mean square error (MMSE) estimation. In particular, we consider the problem of estimating the input signal from the output of an additive white Gaussian channel whose gain is fixed, but unknown. The input distribution is known, and the estimation process consists of two algorithms. First, a channel estimator blindly estimates the channel gain using past observations. Second, a mismatched MMSE estimator, optimized for the estimated channel gain, estimates the input signal. We analyze the regret, i.e., the additional mean square error, that is raised in this process. We derive upper-bounds on both absolute and relative regrets. Bounds are expressed in terms of the Fisher information. We also study regret for unbiased, efficient channel estimators, and derive a simple trade-off between Fisher information and relative regret. This trade-off shows that the product of a certain function of relative regret and Fisher information equals the signal-to-noise ratio, independent of the input distribution. The trade-off relation implies that higher Fisher information results to smaller expected relative regret.

Massively parallel processing of signals in dense microphone arrays

Arrays with large number of microphones can be very effective on audio processing tasks, like denoising, acoustic echo removal, etc. New microphone technologies enable creating large arrays with very low cost per component, but the system can still be very expensive due to costs of transmitting all signals to a single processor, and the computational resources to process the large amount of data. We show how a massively-parallel signal processing approach can solve the cost issues, when applied to the problem of sound source localization. We consider the case where each microphone is coupled to simple processing circuitry, which have full-bandwidth access to data from a few other microphones, while only shared power and low-bandwidth connections are provided between each microphone and a central processor. We discuss implementation issues, and show experimental results obtained in simulations and in microphone array measurements.

Performance analysis for blind identification of acoustic channels

Multichannel blind system identification is a prominent part in audio dereverberation. Despite much progress in this field, the performance of existing algorithms is still unsatisfactory and existing theories have failed to properly explain the issue. To assess performance impediments, we introduce, quantify, and illustrate three major sources of errors: namely, estimation, approximation, and referencing errors. For each error, we provide simple expressions that describe the effect of various system parameters and that can serve as a guideline to explain reality and improve performance.

Decision-Making with Unbounded Loss Functions

We consider the problem of decision-making under uncertainty with unbounded loss functions. Inspired by PAC learning model, we use a slightly different model that incorporates the notion of side information in a more generic form to make it applicable to a broader class of applications including system identification and parameter estimation. We address sufficient conditions for consistent decision-making as well as exponential convergence behavior. In this regard, besides a requirement on the growth function of the class of loss functions, it suffices to have a dominating function whose Orlicz expectation is uniformly bounded over the probabilistic model. Decay exponent, decay rate, and sample complexity for expected risk minimization decision policy are discussed, as well

A subsequence-histogram method for generic vocabulary recognition over deletion channels

We consider the problem of recognizing a vocabulary-a collection of words (sequences) over a finite alphabet-from a potential subsequence of one of its words. We assume the given subsequence is received through a deletion channel as a result of transmission of a random word from one of the two generic underlying vocabularies. An exact maximum a posterior (MAP) solution for this problem counts the number of ways a given subsequence can be derived from particular subsets of candidate vocabularies, requiring exponential time or space.

On Regularization of Least Square Problems via Quadratic Constraints

We consider uncertainty reduction in least square problems raised in system identification with unknown state space. We assume existence of some prior information obtained through a finite series of measurements. This data is modeled in the form of a finite collection of quadratic constraints enclosing the state space. A simple closed form expression is derived for the optimal solution featuring geometric insights and intuitions that reveal a two-fold effort in reducing uncertainty: by correcting the observation error and by improving the condition number of the data matrix. To deal with the dual problem of finding the optimal Lagrange multipliers, we introduce an approximate, positive semidefinite program that can be easily solved using the standard numerical techniques.

System Identification with Unbounded Loss Functions Under Algorithmic Deficiency

We describe and analyze a comprehensive learning model to address issues such as consistency, convergence rate, and sample complexity in the general context of system identification. The learning model is based on unbounded loss functions, and it incorporates a measure of algorithmic deficiency. We define and use a novel formulation of algorithmic solution that is an extension of the empirical risk minimization method in the sense that it uses a generic notion of side information as opposed to the commonly used input/output observation of a system. Sufficient conditions for consistency as well as closed form expressions for exponential convergence rate and sample complexity of the identification algorithm are derived