Computational speech segregation based on an auditory-inspired modulation analysis

The benefit of combining a deep neural network architecture with ideal ratio mask estimation in computational speech segregation to improve speech intelligibility

Computational speech segregation attempts to automatically separate speech from noise. This is challenging in conditions with interfering talkers and low signal-to-noise ratios. Recent approaches have adopted deep neural networks and successfully demonstrated speech intelligibility improvements. A selection of components may be responsible for the success with these state-of-the-art approaches: the system architecture, a time frame concatenation technique and the learning objective. The aim of this study was to explore the roles and the relative contributions of these components by measuring speech intelligibility in normal-hearing listeners. A substantial improvement of 25.4 percentage points in speech intelligibility scores was found going from a subband-based architecture, in which a Gaussian Mixture Model-based classifier predicts the distributions of speech and noise for each frequency channel, to a state-of-the-art deep neural network-based architecture. Another improvement of 13.9 percentage points was obtained by changing the learning objective from the ideal binary mask, in which individual time-frequency units are labeled as either speech- or noise-dominated, to the ideal ratio mask, where the units are assigned a continuous value between zero and one. Therefore, both components play significant roles and by combining them, speech intelligibility improvements were obtained in a six-talker condition at a low signal-to-noise ratio.

The impact of exploiting spectro-temporal context in computational speech segregation

Computational speech segregation aims to automatically segregate speech from interfering noise, often by employing ideal binary mask estimation. Several studies have tried to exploit contextual information in speech to improve mask estimation accuracy by using two frequently-used strategies that (1) incorporate delta features and (2) employ support vector machine (SVM) based integration. In this study, two experiments were conducted. In Experiment I, the impact of exploiting spectro-temporal context using these strategies was investigated in stationary and six-talker noise. In Experiment II, the delta features were explored in detail and tested in a setup that considered novel noise segments of the six-talker noise. Computing delta features led to higher intelligibility than employing SVM based integration and intelligibility increased with the amount of spectral information exploited via the delta features. The system did not, however, generalize well to novel segments of this noise type. Measured intelligibility was subsequently compared to extended short-term objective intelligibility, hit-false alarm rate, and the amount of mask clustering. None of these objective measures alone could account for measured intelligibility. The findings may have implications for the design of speech segregation systems, and for the selection of a cost function that correlates with intelligibility.

Outcome measures based on classification performance fail to predict the intelligibility of binary-masked speech

To date, the most commonly used outcome measure for assessing ideal binary mask estimation algorithms is based on the difference between the hit rate and the false alarm rate (H-FA). Recently, the error distribution has been shown to substantially affect intelligibility. However, H-FA treats each mask unit independently and does not take into account how errors are distributed. Alternatively, algorithms can be evaluated with the short-time objective intelligibility (STOI) metric using the reconstructed speech. This study investigates the ability of H-FA and STOI to predict intelligibility for binary-masked speech using masks with different error distributions. The results demonstrate the inability of H-FA to predict the behavioral intelligibility and also illustrate the limitations of STOI. Since every estimation algorithm will make errors that are distributed in different ways, performance evaluations should not be made solely on the basis of these metrics.

A Deep Ensemble Learning Method for Monaural Speech Separation

Monaural speech separation is a fundamental problem in robust speech processing. Recently, deep neural network (DNN)-based speech separation methods, which predict either clean speech or an ideal time-frequency mask, have demonstrated remarkable performance improvement. However, a single DNN with a given window length does not leverage contextual information sufficiently, and the differences between the two optimization objectives are not well understood. In this paper, we propose a deep ensemble method, named multicontext networks, to address monaural speech separation. The first multicontext network averages the outputs of multiple DNNs whose inputs employ different window lengths. The second multicontext network is a stack of multiple DNNs. Each DNN in a module of the stack takes the concatenation of original acoustic features and expansion of the soft output of the lower module as its input, and predicts the ratio mask of the target speaker; the DNNs in the same module employ different contexts. We have conducted extensive experiments with three speech corpora. The results demonstrate the effectiveness of the proposed method. We have also compared the two optimization objectives systematically and found that predicting the ideal time-frequency mask is more efficient in utilizing clean training speech, while predicting clean speech is less sensitive to SNR variations.

Cochlear implant speech intelligibility outcomes with structured and unstructured binary mask errors

It has been shown that intelligibility can be improved for cochlear implant (CI) recipients with the ideal binary mask (IBM). In realistic scenarios where prior information is unavailable, however, the IBM must be estimated, and these estimations will inevitably contain errors. Although the effects of both unstructured and structured binary mask errors have been investigated with normal-hearing (NH) listeners, they have not been investigated with CI recipients. This study assesses these effects with CI recipients using masks that have been generated systematically with a statistical model. The results demonstrate that clustering of mask errors substantially decreases the tolerance of errors, that incorrectly removing target-dominated regions can be as detrimental to intelligibility as incorrectly adding interferer-dominated regions, and that the individual tolerances of the different types of errors can change when both are present. These trends follow those of NH listeners. However, analysis with a mixed effects model suggests that CI recipients tend to be less tolerant than NH listeners to mask errors in most conditions, at least with respect to the testing methods in each of the studies. This study clearly demonstrates that structure influences the tolerance of errors and therefore should be considered when analyzing binary-masking algorithms.