Structure in time-frequency binary masking errors and its impact on speech intelligibility

Abstract

Although requiring prior knowledge makes the ideal binary mask an impractical algorithm, substantial increases in measured intelligibility make it a desirable benchmark. While this benchmark has been studied extensively, many questions remain about the factors that influence the intelligibility of binary-masked speech with non-ideal masks. To date, researchers have used primarily uniformly random, uncorrelated mask errors and independently presented error types (i.e., false positives and negatives) to characterize the influence of estimation errors on intelligibility. However, practical estimation algorithms produce masks that contain errors of both types and with non-trivial amounts of structure. This paper introduces an investigation framework for binary masks and presents listener studies that use this framework to illustrate how interactions between error types and structure affect intelligibility. First, this study demonstrates that clustering (i.e., a form of structure) of mask errors reduces intelligibility. Furthermore, while previous research has suggested that false positives are more detrimental to intelligibility than false negatives, this study indicates that false negatives can be equally detrimental to intelligibility when they contain structure or when both error types are present. Finally, this study shows that listeners tolerate fewer mask errors when both types of errors are present, especially when the errors contain structure.

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

“…However, H-FA fails to predict the impact of the distribution of errors, and instead, predicts that all masks with the same error rates yield the same intelligibility outcome. Thus, even though the correlation between mean H-FA and behavioral scores for conditions with c ¼ 2.0 (i.e., the conditions with an error distribution that most closely match the error distribution of the estimated masks of Kim et al;Kressner and Rozell, 2015) is high (r ¼ 0.97), H-FA is unable to account for the differences in the behavioral scores that arise when masks contain errors that are distributed differently. In contrast to H-FA, STOI is able to qualitatively predict the trends in the behavioral data when false negative errors are presented, as demonstrated by the similarities between Figs.…”

“…RESULTS Figure 1 shows the behavioral results and predicted scores for the first two experiments of Kressner and Rozell (2015). Figure 1(a) shows the behavioral results when false positive errors are introduced (i.e., more energy from the interferer-dominated T-F units is erroneously retained), whereas Fig.…”

“…The masks and mixture signals were regenerated and processed for each of the three experiments using the same procedures as in Kressner and Rozell (2015). Then H-FA and STOI were computed for each individual sentence.…”

“…The objective measures H-FA and STOI were assessed on their ability to predict the intelligibility scores from the listener studies in Kressner and Rozell (2015). In these experiments, speech mixed with babble was processed with binary masks generated from a statistical model that artificially introduced errors with parametrically controlled distributions.…”

“…To investigate the influence of different distributions of these errors on speech intelligibility outcomes, Kressner and Rozell (2015) and Kressner et al (2016) scored normal hearing (NH) listeners and cochlear implant (CI) recipients, respectively, on their word recognition of noisy speech that had been processed with binary masks containing different distributions of errors. These studies together demonstrate that the impact of false positive and false negative error rates on speech intelligibility scores is highly dependent on how the errors are distributed.…”