Frederic Mustiere

A Modified Rao-Blackwellised Particle Filter

Rao-Blackwellised particle filters (RBPFs) are a class of particle filters (PFs) that exploit conditional dependencies between parts of the state to estimate. By doing so, RBPFs can improve the estimation quality while also reducing the over-all computational load in comparison to original PFs. However, the computational complexity is still too high for many real-time applications. In this paper, we propose a modified RBPF that requires a single Kalman Filter (KF) iteration per input sample. Comparative experiments show that while good convergence can still be obtained, computational efficiency is always drastically increased, making this algorithm an option to consider for real-time implementations

Rao-Blackwellised Particle Filters: Examples of Applications

In this work, we present some examples of applications of the so-called Rao-Blackwellised particle filter (RBPF). RBPFs are an extension to particle filters (PFs) which are applicable to conditionally linear-Gaussian state-space models. Although RBPF introductions and reviews may be found in many existing sources, going through the specific vocabulary and concepts of particle filtering can sometimes prove to be time-consuming for the non-initiated reader willing to experiment with alternative algorithms. The goal of the paper is to introduce RBPF-based methods in an accessible manner via a main algorithm, which is detailed enough to be readily applied to a wide range of problems. To illustrate the practicality and the convenience of the approach, the algorithm is then tailored to two examples from different fields. The first example is related to system identification, and the second is an application of speech enhancement

Low-cost modifications of Rao-Blackwellized particle filters for improved speech denoising

Rao-Blackwellized particle filtering (RBPF) has established itself as an appealing alternative to classical Kalman filter-based speech enhancement schemes. Its advantages include a high noise reduction between utterances, and flexibility in the definition of model parameters. There is, however, still room for improvement in reducing the level of residual noise within the estimated speech. In this paper, low-cost ways to improve the quality of enhanced speech are proposed, starting from an existing RBPF algorithm. In the first method, the RBPF is slightly modified to produce improved but delayed estimates. The next two methods are centered on the inclusion of low-order FIR filters during or after the estimation process. Finally, a simple way to combine different speech estimates is shown. The main interest is that even rough estimates obtained with lightweight algorithms can be used to improve the resulting quality. The idea is centered on a single Kalman filter running on mixing proportions. Various simulation results are shown and rated using their overall SNR, their segmental SNR, and their PESQ (wideband extension) score. Advantages and drawbacks of each method are highlighted. Overall, the changes proposed are found to be beneficial across most or all of the measures used.

Fast communication: Design of multichannel frequency domain statistical-based enhancement systems preserving spatial cues via spectral distances minimization

It is often very important for multichannel speech enhancement systems, such as hearing aids, to preserve spatial impressions. Usually, this is achieved by first designing a particular speech enhancement algorithm and later or separately constraining the obtained solution to respect spatial cues. Instead, we propose in this paper to conduct the entire systems design via the minimization of statistical spectral distances seen as functions of a real-valued, common gain to be applied to all channels in the frequency-domain. For various spectral distances, we show that the gain derived is expressible in terms of optimal multichannel spectral amplitude estimators (such as the multichannel Minimum Mean Squared Error Spectral Amplitude Estimator, among others). In addition, we report experimental results in complex environments (i.e., including reverberation, interfering talkers, and low signal-to-noise ratio), showing the potential of the proposed methods against recent state-of-the-art multichannel enhancement setups which preserve spatial cues as well.

Improved colored noise handling in Kalman Filter-based speech enhancement algorithms

This paper presents a simple alternative to the traditional handling of autoregressive colored observation noise processes in Kalman filter-based speech enhancement algorithms. The method is entirely centered on a rewriting of the state-space equations describing the problem. The proposed approach decreases the dimension of the state vector and the amount of computations per iteration, and also naturally reduces to the white noise case when a zero-order autoregressive colored noise is chosen. In addition, from the multiple experiments conducted using several Kalman filter-based algorithms, it is found that the quality obtained with the new method, as measured by different speech quality measures, is equivalent and in some cases better. The simulations presented are based on both computer-generated and real-world colored noises, in stationary and nonstationary cases.

Speech Enhancement Based on Nonlinear Models Using Particle Filters

Motivated by the reportedly strong performance of particle filters (PFs) for noise reduction on essentially linear speech production models, and the mounting evidence that the introduction of nonlinearities can lead to a refined speech model, this paper presents a study of PF solutions to the problem of speech enhancement in the context of nonlinear, neural-type speech models. Several variations of a global model are presented (single/multiple neurons; bias/no bias), and corresponding PF solutions are derived. Different importance functions are given when beneficial, Rao-Blackwellization is proposed when possible, and dual/nondual versions of each algorithms are presented. The method shown can handle both white and colored noise. Using a variety of speech and noise signals and different objective quality measures, the performance of these algorithms are evaluated against other PF solutions running on linear models, as well as some traditional enhancement algorithms. A certain hierarchy in performance is established between each algorithm in the paper. Depending on the experimental conditions, the best-performing algorithms are a classical Rao-Blackwellized particle filter (RBPF) running on a linear model, and a proposed PF employing a nondual, nonlinear model with multiple neurons and no biases. With consistence, the neural-network-based PF outperforms RBPF at low signal-to-noise ratio (SNR).

Neural-based approach to perceptual sparse coding of audio signals

We propose a neural architecture for the perceptual sparse coding of audio signals based on a previously proposed technique called Local Competitive Algorithm (LCA) that was originally applied to image and video coding. LCAs are designed to be implemented in a dynamical system composed of many neuron-like elements operating in parallel. For the processing of audio signals, we here use gammatone filters that mimic the behavior of the auditory pathway as the receptive field of our neurons. We also adapted LCA to time-varying audio signals. Given the fact that LCA does not take into account the difference between perceived coding error and mathematical Mean-Squared Error (MSE), we propose in this article the Perceptual Local Competitive Algorithm (PLCA) and derive a convergence formula, as well as a corresponding neural architecture for it.We show that our proposed PLCA minimizes the perceptual coding error and can model phenomena such as absolute threshold of hearing and masking. Our sparse audio coder based on PLCA compares with the more conventional greedy (i.e., matching pursuit) algorithms for sparse coding in terms of quality and can be implemented in a much faster way, especially when parallel processing units (i.e., embedded circuits) can be afforded. We also show that our proposed PLCA is much more robust to quantization error than the conventional matching pursuit for audio coding.

Bandwidth extension for speech enhancement

In real-world wideband noisy recordings of speech, the Signal-to-Noise ratios in the lower half spectra are typically higher than those observed in their upper half counterparts, due to the average frequency distribution of the speech energy. Speech enhancement algorithms therefore struggle to recover damaged high frequency components, thereby penalizing the output perceived quality of the enhanced signal. In this paper, it is proposed to use and adapt the concept of Bandwidth Extension in the context of speech enhancement so as to reinforce the high-frequency estimates of the clean speech. By allowing enhancement schemes to focus their resources on narrowband speech while synthesizing the rest of the signal, it is found that the overall quality can be improved, while reducing significantly the computational costs of the obtained method.

Real-world particle filtering-based speech enhancement

This paper presents a viable particle filtering (PF) solution for single microphone speech enhancement in real-world conditions, i.e., operating at low SNR in nonstationary noise environments, while remaining computationally tractable. The enhancement takes place in the subband domain with elementary PFs in each band. To efficiently handle complex noise situations, the noise spectrum is modelled in each band as a white Gaussian noise sequence with a time-varying gain. Two solutions are proposed to estimate these time-varying average subband noise levels: they are either drawn internally by the PFs, or they are obtained by external dedicated noise power spectral density estimation - both methods are found to yield very close results. Several subband decompositions are tested, and a robust way of incorporating perceptual constraining is introduced. The assembled PF-based architecture is then compared with state-of-the-art enhancement algorithms in various conditions, and is found to outperform them according to seven objective speech quality measures.

A fast convergence two-step procedure for AR modeling of power spectral densities

A new technique for the minimization of customary cost functions for all-pole modeling of power spectral densities is presented. In the literature, optimizations are usually based on unnormalized autoregressive (AR) coefficients. In contrast, the proposed method is centered on a two-step descent using normalized AR coefficients on the one hand and the residual power on the other hand. For each cost function, efficient ways to obtain gradients are derived and a descent-based optimization is formulated. The resulting procedure converges significantly faster than the corresponding gradient descents on un-normalized coefficients, while still being computationally efficient. In addition to the traditional Yule-Walker distance, the Itakura-Saito distance, the COSH distance, the RMS log-spectral ratio distance, and a mean-squared error cost function are treated. Convergence results and curves are presented accordingly for different situations.