Eugen Hoffmann

Independent Component Analysis and Time-Frequency Masking for Speech Recognition in Multitalker Conditions

When a number of speakers are simultaneously active, for example in meetings or noisy public places, the sources of interest need to be separated from interfering speakers and from each other in order to be robustly recognized. Independent component analysis (ICA) has proven a valuable tool for this purpose. However, ICA outputs can still contain strong residual components of the interfering speakers whenever noise or reverberation is high. In such cases, nonlinear postprocessing can be applied to the ICA outputs, for the purpose of reducing remaining interferences. In order to improve robustness to the artefacts and loss of information caused by this process, recognition can be greatly enhanced by considering the processed speech feature vector as a random variable with time-varying uncertainty, rather than as deterministic. The aim of this paper is to show the potential to improve recognition of multiple overlapping speech signals through nonlinear postprocessing together with uncertainty-based decoding techniques.

A batch algorithm for blind source separation of acoustic signals using ICA and time-frequency masking

The problem of Blind Source Separation (BSS) of convolved acoustic signals is of great interest for many classes of applications such as in-car speech recognition, hands-free telephony or hearing devices. Due to the convolutive mixing process, the source separation is performed in the frequency domain, using Independent Component Analysis (ICA). However the quality of solution of the ICA-algorithms can be improved by applying time-frequency masking. In this paper we present a batch-algorithm for time-frequency masking using the time-frequency structure of separated signals.

Using information theoretic distance measures for solving the permutation problem of blind source separation of speech signals

The problem of blind source separation (BSS) of convolved acoustic signals is of great interest for many classes of applications. Due to the convolutive mixing process, the source separation is performed in the frequency domain, using independent component analysis (ICA). However, frequency domain BSS involves several major problems that must be solved. One of these is the permutation problem. The permutation ambiguity of ICA needs to be resolved so that each separated signal contains the frequency components of only one source signal. This article presents a class of methods for solving the permutation problem based on information theoretic distance measures. The proposed algorithms have been tested on different real-room speech mixtures with different reverberation times in conjunction with different ICA algorithms.

A Soft Masking Strategy Based on Multichannel Speech Probability Estimation for Source Separation and Robust Speech Recognition

In this paper, we present a post processing algorithm that improves the quality of the results of ICA-algorithms by applying a modified speech enhancement technique. The proposed method is based on estimating speech probabilities from the ICA outputs by means of two dimensional correlations. With these probabilities, a soft masking function can be applied on the ICA outputs, which results in significantly increased interferer suppression. In order to avoid negative influences on subsequent speech recognition, missing feature recognition has been applied to robustly recognize the non-linearly processed speech signal. The algorithm has been tested on real-room speech mixtures with a reverberation time of 300ms, where an SIR-improvement of up to 32dB has been obtained, which was 10dB above ICA performance for the same dataset.

Recognition of Multiple Speech Sources Using ICA

In meetings or noisy public places, often a number of speakers are active simultaneously and the sources of interest need to be separated from interfering speech in order to be robustly recognized. Independent component analysis (ICA) has proven to be a valuable tool for this purpose. However, under difficult environmental conditions, ICA outputs may still contain strong residual components of the interfering speakers. In such cases, time-frequency masking can be applied to the ICA outputs to reduce the remaining interferences. In order to remain robust against possible resulting artifacts and loss of information, treating the processed speech feature vector as a random variable with time-varying uncertainty, rather than as deterministic, is a helpful strategy. This chapter shows the ways of improving recognition of multiple speech signals based on nonlinear postprocessing, applied together with uncertainty-based decoding techniques.

Speech enhancement for automatic speech recognition using complex gaussian mixture priors for noise and speech

Statistical speech enhancement methods often rely on a set of assumptions, like gaussianity of speech and noise processes or perfect knowledge of their parameters, that are not fully met in reality. Recent advancements have shown the potential improvement in speech enhancement obtained by employing supergaussian speech models conditioned on the estimated signal to noise ratio. In this paper we derive a supergaussian model for speech enhancement in which both speech and noise priors are assumed to be complex Gaussian mixture models. We introduce as well a method for the computation of the noise prior based on the noise variance estimator used. Finally, we compare the developed estimators with the conventional Ephraim-Malah filters in the context of robust automatic speech recognition.