Target Speech Research Articles

Domain-specific hearing-in-noise performance is associated with absolute pitch proficiency

Recent evidence suggests that musicians may have an advantage over non-musicians in perceiving speech against noisy backgrounds. Previously, musicians have been compared as a homogenous group, despite demonstrated heterogeneity, which may contribute to discrepancies between studies. Here, we investigated whether “quasi”-absolute pitch (AP) proficiency, viewed as a general trait that varies across a spectrum, accounts for the musician advantage in hearing-in-noise (HIN) performance, irrespective of whether the streams are speech or musical sounds. A cohort of 12 non-musicians and 42 trained musicians stratified into high, medium, or low AP proficiency identified speech or melody targets masked in noise (speech-shaped, multi-talker, and multi-music) under four signal-to-noise ratios (0, − 3, − 6, and − 9 dB). Cognitive abilities associated with HIN benefits, including auditory working memory and use of visuo-spatial cues, were assessed. AP proficiency was verified against pitch adjustment and relative pitch tasks. We found a domain-specific effect on HIN perception: quasi-AP abilities were related to improved perception of melody but not speech targets in noise. The quasi-AP advantage extended to tonal working memory and the use of spatial cues, but only during melodic stream segregation. Overall, the results do not support the putative musician advantage in speech-in-noise perception, but suggest a quasi-AP advantage in perceiving music under noisy environments.

Deficient-basis-complementary rank-constrained spatial covariance matrix estimation based on multivariate generalized Gaussian distribution for blind speech extraction

Rank-constrained spatial covariance matrix estimation (RCSCME) is a blind speech extraction method utilized under the condition that one-directional target speech and diffuse background noise are mixed. In this paper, we propose a new model extension of RCSCME. RCSCME simultaneously conducts both the deficient rank-1 component complementation of the diffuse noise spatial covariance matrix, which is incompletely estimated by preprocessing methods such as independent low-rank matrix analysis, and the estimation of the source model parameters. In the conventional RCSCME, between the two parameters constituting the deficient rank-1 component, only the scale is estimated, whereas the other parameter, the deficient basis, is fixed in advance; however, how to choose the fixed deficient basis is not unique. In the proposed RCSCME model, we also regard the deficient basis as a parameter to estimate. As the generative model of an observed signal, we utilized the super-Gaussian generalized Gaussian distribution, which achieves better separation performance than the Gaussian distribution in the conventional RCSCME. Assuming the model, we derive new majorization-minimization (MM)- and majorization-equalization (ME)-algorithm-based update rules for the deficient basis. In particular, among innumerable ME-algorithm-based update rules, we successfully find an ME-algorithm-based update rule with a mathematical proof supporting the fact that the step of the update rule is larger than that of the MM-algorithm-based update rule. We confirm that the proposed method outperforms conventional methods under several simulated noise conditions and a real noise condition.

Speech understanding in diffuse steady noise in typically hearing and hard of hearing listeners.

Spatial cues can facilitate segregation of target speech from maskers. However, in clinical practice, masked speech understanding is most often evaluated using co-located speech and maskers (i.e., without spatial cues). Many hearing aid centers in France are equipped with five-loudspeaker arrays, allowing masked speech understanding to be measured with spatial cues. It is unclear how hearing status may affect utilization of spatial cues to segregate speech and noise. In this study, speech reception thresholds (SRTs) for target speech in “diffuse noise” (target speech from 1 speaker, noise from the remaining 4 speakers) in 297 adult listeners across 9 Audilab hearing centers. Participants were categorized according to pure-tone-average (PTA) thresholds: typically-hearing (TH; ≤ 20 dB HL), mild hearing loss (Mild; >20 ≤ 40 dB HL), moderate hearing loss 1 (Mod-1; >40 ≤ 55 dB HL), and moderate hearing loss 2 (Mod-2; >55 ≤ 65 dB HL). All participants were tested without aided hearing. SRTs in diffuse noise were significantly correlated with PTA thresholds, age at testing, as well as word and phoneme recognition scores in quiet. Stepwise linear regression analysis showed that SRTs in diffuse noise were significantly predicted by a combination of PTA threshold and word recognition scores in quiet. SRTs were also measured in co-located and diffuse noise in 65 additional participants. SRTs were significantly lower in diffuse noise than in co-located noise only for the TH and Mild groups; masking release with diffuse noise (relative to co-located noise) was significant only for the TH group. The results are consistent with previous studies that found that hard of hearing listeners have greater difficulty using spatial cues to segregate competing speech. The data suggest that speech understanding in diffuse noise provides additional insight into difficulties that hard of hearing individuals experience in complex listening environments.

Affective Political Polarization and Hate Speech: Made for Each Other?

Abundant academic research has shown evidence of the growing polarization across the globe both in general and in terms of affective polarization. Previous research on this topic primarily employed reactive research methods like surveys or experiments, which however do not allow researchers to observe the behavior of the units of analysis in a natural setting. Presents an alternative approach that involves analyzing the observed behavior of social media users and identifying the key polarizing cleavages through the study of hate speech with respect to distinct target groups. We present a novel coding schema for textual data, which includes two components: first, an operationalized definition of hate speech as a phenomenon with at least one of the three elements - insult, discrimination, or aggression; and second, an original coding guide for human coders annotating the use of hate speech. We apply our approach to the analysis of empirical data that includes over 5000 posts on the social media platform VK about the meetings between the Presidents of Russia and Belarus in 2020-2021. After coding the collected data, we performed the empirical analysis that identified two generic cleavages. One is about domestic politics in Belarus and Russia, whereas the other is related to the opposition between these two countries on the one hand, and Western countries on the other. We also found an additional Russian/Belarusian cleavage that is peculiar to the collected dataset. Our methodology also allowed us to identify and analyze the dynamics of macro-groups that were targets of hate speech. Importantly, these results - as any other dynamic aspect of analysis - would be highly challenging in research based on reactive methods. Thereby our results highlight the prospects of applying the proposed methodology to a broad range of textual data, as well as the benefits of exploratory analysis that helps overcome the limitations of survey instruments.

Adapting Multiple Distributions for Bridging Emotions from Different Speech Corpora.

In this paper, we focus on a challenging, but interesting, task in speech emotion recognition (SER), i.e., cross-corpus SER. Unlike conventional SER, a feature distribution mismatch may exist between the labeled source (training) and target (testing) speech samples in cross-corpus SER because they come from different speech emotion corpora, which degrades the performance of most well-performing SER methods. To address this issue, we propose a novel transfer subspace learning method called multiple distribution-adapted regression (MDAR) to bridge the gap between speech samples from different corpora. Specifically, MDAR aims to learn a projection matrix to build the relationship between the source speech features and emotion labels. A novel regularization term called multiple distribution adaption (MDA), consisting of a marginal and two conditional distribution-adapted operations, is designed to collaboratively enable such a discriminative projection matrix to be applicable to the target speech samples, regardless of speech corpus variance. Consequently, by resorting to the learned projection matrix, we are able to predict the emotion labels of target speech samples when only the source label information is given. To evaluate the proposed MDAR method, extensive cross-corpus SER tasks based on three different speech emotion corpora, i.e., EmoDB, eNTERFACE, and CASIA, were designed. Experimental results showed that the proposed MDAR outperformed most recent state-of-the-art transfer subspace learning methods and even performed better than several well-performing deep transfer learning methods in dealing with cross-corpus SER tasks.

Hate speech or free speech: an ethical dilemma?

ABSTRACT Freedom of expression is the cornerstone and the warrant of democracy, but like all other rights and freedoms we enjoy, also the right to free speech has its limits. Hate speech is one of the most resilient manifestations of cyberviolence, and is not to be equalled with free speech. This is partly related to our perception of the freedom of expression that has substantially changed through time; from the past focus on being able to freely express your opinion, to question, dissent and challenge the government, to today’s need to limit these rights in order to protect the targets of hate speech. The article first looks at how hate speech is defined at the level of the EU and at the problems connected with its regulation. It continues with an overview of relevant literature and research on the topic, and concludes by analyzing and interpreting the results of an online survey conducted among Slovenian university students in February 2022. Our goal is to understand the reasons behind the rise of hate speech and reactions to it; from its acceptability and underestimation to indifference and ignorance.

Auditory motion as a cue for source segregation and selection in a "cocktail party" listening environment.

Source motion was examined as a cue for segregating concurrent speech or noise sources. In two different headphone-based tasks-motion detection (MD) and speech-on-speech masking (SI)-one source among three was designated as the target only by imposing sinusoidal variation in azimuth during the stimulus presentation. For MD, the lstener was asked which of the three concurrent sources was in motion during the trial. For SI, the listener was asked to report the words spoken by the moving speech source. MD performance improved as the amplitude of the sinusoidal motion (i.e., displacement in azimuth) increased over the range of values tested (±5° to ±30°) for both modulated noise and speech targets, with better performance found for speech. SI performance also improved as the amplitude of target motion increased. Furthermore, SI performance improved as word position progressed throughout the sentence. Performance on the MD task was correlated with performance on SI task across individual subjects. For the SI conditions tested here, these findings are consistent with the proposition that listeners first detect the moving target source, then focus attention on the target location as the target sentence unfolds.

Auditory spatial analysis in reverberant multi-talker environments with congruent and incongruent audio-visual room information.

In a multi-talker situation, listeners have the challenge of identifying a target speech source out of a mixture of interfering background noises. In the current study, it was investigated how listeners analyze audio-visual scenes with varying complexity in terms of number of talkers and reverberation. The visual information of the room was either congruent with the acoustic room or incongruent. The listeners' task was to locate an ongoing speech source in a mixture of other speech sources. The three-dimensional audio-visual scenarios were presented using a loudspeaker array and virtual reality glasses. It was shown that room reverberation, as well as the number of talkers in a scene, influence the ability to analyze an auditory scene in terms of accuracy and response time. Incongruent visual information of the room did not affect this ability. When few talkers were presented simultaneously, listeners were able to detect a target talker quickly and accurately even in adverse room acoustical conditions. Reverberation started to affect the response time when four or more talkers were presented. The number of talkers became a significant factor for five or more simultaneous talkers.

Spatial hearing in middle-aged and older adults with a hearing aid: Does noise type matter?

Purpose: The present study aimed to identify the overall patterns of spatial hearing in a sample population of the geriatric population wearing a hearing aid in an actual clinical setting and to investigate while paying attention to aging and the effects of spatial separation on informational masking (IM) of target speech of common types of noise in everyday life.Methods: Nine participants (mean age: 68.9 ± 9.6) wearing hearing aids participated. The spatial hearing abilities of the participants were assessed by measuring their speech reception threshold (SRT) and spatial release from masking (SRM), using speech-like (SLN) and steady-state (SSN) noise-maskers.Results: The study results identified a trend in the decrease in SRT in all noise conditions with increasing azimuth angles. The larger the target/masker spatial separation, the larger the SRM became in all noise conditions. Results revealed that geriatric hearing-impaired participants experienced a comparable amount of SRM from IM in SLN concerning SSN.Conclusions: Given these findings, the spatial separation did not have the same effects on the IM of target speech by SLN in the geriatric population wearing a hearing aid as it did in the general hearing-impaired population at younger ages. It can be inferred from these findings that the extent to which informational maskers can compromise the spatial hearing abilities of geriatric hearing-impaired listeners may depend on their age and the severity of spatial hearing loss being imposed on them.

Implicitly Aligning Joint Distributions for Cross-Corpus Speech Emotion Recognition

In this paper, we investigate the problem of cross-corpus speech emotion recognition (SER), in which the training (source) and testing (target) speech samples belong to different corpora. This case thus leads to a feature distribution mismatch between the source and target speech samples. Hence, the performance of most existing SER methods drops sharply. To solve this problem, we propose a simple yet effective transfer subspace learning method called joint distribution implicitly aligned subspace learning (JIASL). The basic idea of JIASL is very straightforward, i.e., building an emotion discriminative and corpus invariant linear regression model under an implicit distribution alignment strategy. Following this idea, we first make use of the source speech features and emotion labels to endow such a regression model with emotion-discriminative ability. Then, a well-designed reconstruction regularization term, jointly considering the marginal and conditional distribution alignments between the speech samples in both corpora, is adopted to implicitly enable the regression model to predict the emotion labels of target speech samples. To evaluate the performance of our proposed JIASL, extensive cross-corpus SER experiments are carried out, and the results demonstrate the promising performance of the proposed JIASL in coping with the tasks of cross-corpus SER.

Subjective tinnitus patients with normal pure-tone hearing still suffer more informational masking in the noisy environment.

Subjective tinnitus patients experience more hearing difficulties than normal peers in complex hearing environments, even though most of these patients have normal pure-tone hearing thresholds. Using speech recognition tasks under different masking conditions can provide insight into whether the effects of tinnitus are lateralized and the mechanisms behind the effects. By simulating sound field recordings, we obtain a target speech sentence that can be perceived as presented on one side and noise or speech masking with or without spatial separation from it. Our study used the virtual sound field technique to investigate the difference in speech recognition ability between chronic subjective tinnitus patients and a normal-hearing control group under the four masking conditions (speech-spectrum noise masking or two-talker speech masking, with or without perceived spatial separation). Experiment 1 showed no differences for target speech perceived location (left or right), which rules out a lateralization of the effect of tinnitus patients. Experiment 2 further found that although tinnitus patients had weaker performance than normal people in very complex auditory scenarios, when the spatial cue of the target speech exists, they can make good use of this cue to make up for the original processing disadvantage and achieve a similar performance as the normal-hearing group. In addition, the current study distinguished the effects of informational masking and energetic masking on speech recognition in patients with tinnitus and normal hearing. The results suggest that the impact of tinnitus on speech recognition in patients is more likely to occur in the auditory center rather than the periphery.

Target Speaker Extraction by Fusing Voiceprint Features

It is a critical problem to accurately separate clean speech in the multispeaker scenario for different speakers. However, in most cases, smart devices such as smart phones interact with only one specific user. As a consequence, the speech separation models adopted by these devices only have to extract the target speaker’s speech. A voiceprint, which reflects the speaker’s voice characteristics, provides prior knowledge for the target speech separation. Therefore, how to efficiently integrate voiceprint features into the existing speech separation models to improve their performance for the target speech separation is an interesting problem not fully explored. This paper attempts to solve this issue to some extent and our contributions are as follows. First, two different voiceprint features (i.e., MFCCs and d-vector) are explored in the performance enhancement for three speech separation models. Second, three different feature fusion methods are proposed to efficiently fuse the voiceprint features with the magnitude spectrograms originally used in the speech separation models. Third, a target speech extraction method which utilizes the fused features is proposed for two speaker-independent models. Experiments demonstrate that the speech separation models integrated with voiceprint features using three feature fusion methods can effectively extract the target speaker’s speech.

Coping and adjustment during COVID-19

While the initial outbreak of COVID-19 in China impacted Chinese international students’ families’ lives, the spread of the virus in the U.S. heavily influenced their own. Drawing upon the stress and coping model (Lazarus & Folkman, 1984), this qualitative study unpacks Chinese doctoral students’ experiences and coping at one large research university in the U.S. during the pandemic. The open-ended interviews with eight Chinese international doctoral students revealed the stressors associated with the sudden changes in participants’ personal, social and academic lives. Findings showed that participants experienced learning obstacles, health concerns, funding uncertainties, and limited social interactions. Participants mainly utilized emotion-focused strategies to cope with the daily life stressors, the tense political climate and hate speech targeting Chinese people in the U.S. This study contributes to the dialogue about stress coping in the pandemic and suggests education practitioners possible improvements in student services.

Age-Related Differences in Early Cortical Representations of Target Speech Masked by Either Steady-State Noise or Competing Speech.

Word in noise identification is facilitated by acoustic differences between target and competing sounds and temporal separation between the onset of the masker and that of the target. Younger and older adults are able to take advantage of onset delay when the masker is dissimilar (Noise) to the target word, but only younger adults are able to do so when the masker is similar (Babble). We examined the neural underpinning of this age difference using cortical evoked responses to words masked by either Babble or Noise when the masker preceded the target word by 100 or 600 ms in younger and older adults, after adjusting the signal-to-noise ratios (SNRs) to equate behavioural performance across age groups and conditions. For the 100 ms onset delay, the word in noise elicited an acoustic change complex (ACC) response that was comparable in younger and older adults. For the 600 ms onset delay, the ACC was modulated by both masker type and age. In older adults, the ACC to a word in babble was not affected by the increase in onset delay whereas younger adults showed a benefit from longer delays. Hence, the age difference in sensitivity to temporal delay is indexed by early activity in the auditory cortex. These results are consistent with the hypothesis that an increase in onset delay improves stream segregation in younger adults in both noise and babble, but only in noise for older adults and that this change in stream segregation is evident in early cortical processes.

Effects of reverberation on speech intelligibility in noise for hearing-impaired listeners.

Reverberation can have a strong detrimental effect on speech intelligibility in noise. Two main monaural effects were studied here: the temporal smearing of the target speech, which makes the speech less understandable, and the temporal smearing of the noise, which reduces the opportunity for listening in the masker dips. These phenomena have been shown to affect normal-hearing (NH) listeners. The aim of this study was to understand whether hearing-impaired (HI) listeners are more affected by reverberation, and if so to identify which of these two effects is responsible. They were investigated separately and in combination, by applying reverberation either on the target speech, on the noise masker, or on both sources. Binaural effects were not investigated here. Intelligibility scores in the presence of stationary and modulated noise were systematically compared for both NH and HI listeners in these situations. At the optimal signal-to-noise ratios (SNRs) (that is to say, the SNRs with the least amount of floor and ceiling effects), the temporal smearing of both the speech and the noise had a similar effect for the HI and NH listeners, so that reverberation was not more detrimental for the HI listeners. There was only a very limited dip listening benefit at this SNR for either group. Some differences across group appeared at the SNR maximizing dip listening, but they could not be directly related to an effect of reverberation, and were rather due to floor effects or to the reduced ability of the HI listeners to benefit from dip listening, even in the absence of reverberation.

NATIONAL MASS MEDIA PARTIALITY (ANALYSIS OF LANGUAGE USE IN POLITICAL DISCOURSE)

A reliable newspaper should accommodate heterogeneity regarding geographical and psychological readers. The wider the coverage area of a newspaper, the more heterogeneous its readers will be. By putting more concerns on these two dimensions, a newspaper is expected to be in a neutral position. This impartiality or neutrality of the mass media becomes a more intriguing topic when it is associated with the theory arguing that language cannot be neutral since it is a form of social practice, as Halliday has always emphasized. This study aims to describe (1) the use of language (speech acts, speech targets, and speech objectives) in the political discourse contained in the editorials of Republika and Kompas newspapers and (2) the objectivity of the two newspapers. This study used a qualitative-critical approach. This type of research is descriptive-interpretative research and, more specifically, is a type of critical pragmatic analysis research. The analysis of the use of language in the editorials of Koran Republika and Koran Kompas shows that the two newspapers have different tendencies due to different ideologies. The vision and history of a newspaper can determine its ideology of a newspaper. Republika is more inclined to pair 01 (Joko Widodo-Makruf Amin) and share a similar background related to Islam. Kompas is more inclined to pair serial number 02 (Prabowo Subianto-Sandiaga Uno), one of whom is not an Islamic figure.

Masked Speech Recognition as a Function of Masker Location for Cochlear Implant Users With Single-Sided Deafness.

Cochlear implant (CI) recipients with normal or near normal hearing (NH) in the contralateral ear, referred to as single-sided deafness (SSD), experience significantly better speech recognition in noise with their CI than without it, although reported outcomes vary. One possible explanation for differences in outcomes across studies could be differences in the spatial configurations used to assess performance. This study compared speech recognition for different spatial configurations of the target and masker, with test materials used clinically. Sixteen CI users with SSD completed tasks of masked speech recognition presented in five spatial configurations. The target speech was presented from the front speaker (0° azimuth). The masker was located either 90° or 45° toward the CI-ear or NH-ear or colocated with the target. Materials were the AzBio sentences in a 10-talker masker and the Bamford-Kowal-Bench Speech-in-Noise test (BKB-SIN; four-talker masker). Spatial release from masking (SRM) was computed as the benefit associated with spatial separation relative to the colocated condition. Performance was significantly better when the masker was separated toward the CI-ear as compared to colocated. No benefit was observed for spatial separations toward the NH-ear. The magnitude of SRM for spatial separations toward the CI-ear was similar for 45° and 90° when tested with the AzBio sentences, but a larger benefit was observed for 90° as compared to 45° for the BKB-SIN. Masked speech recognition in CI users with SSD varies as a function of the spatial configuration of the target and masker. Results supported an expansion of the clinical test battery at the study site to assess binaural hearing abilities for CI candidates and recipients with SSD. The revised test battery presents the target from the front speaker and the masker colocated with the target, 90° toward the CI-ear, or 90° toward the NH-ear.

Web-Questionnaire-Based Corpus Creation Under Assumption of Human as Speech Targets

This paper presents a method that uses a web questionnaire to create a corpus containing spontaneous utterances of natural ideas, which may contain grammatical mistakes. In an experimental implementation of the method, the subjects were informed that they were receiving nursing care from a person, and they were required to answer a web-based questionnaire in which their responses were recorded as speech utterances. Compared to the Wizard of Oz approach and interview-based corpus-creation methods, the presented method simplifies the collection of utterances. Furthermore, we conducted a two-fold assessment to verify the effectiveness of the presented method. First, the approach exhibited a significant reduction in workload compared to interview-style utterance collection. Second, we compared the variety of expressions collected when subjects were informed that they were talking to a person with those collected when they were informed that they were communicating with a nursing robot. The results indicate that, although the number of utterances was larger for a robot than for a person, in terms of other metrics such as time efficiency index, the total number of morphemes, the average number of morphemes per utterance, the number of unique morphemes, and coefficient of variation, the utterances were larger for a human speech target than for a robot.

Perspectives of Muslim and Minority Canadian Youth on Hate Speech and Social Media

In this article, we highlight the perspectives of marginalized Canadian youth regarding hate speech on social media. Specifically, our research focus is on the complexity and intersectionality involved in cyber violence, especially in relation to marginalized identities. Twenty-five participants aged 18 to 25 studying at a central Canadian University (from an initial sample of 90 participants) who self-identified as victims of hate speech were invited to share their experiences and narrate their stories. Research results demonstrate that online hate speech is growing in Canada to an extent where it is has become normalized. This has serious implications for the well-being of Canadian youth - both perpetrators and victims of hate speech. The main targets of hate speech on social media in Canada are immigrants and minorities, particularly Muslims. Results show that online hate speech has significant consequences for the lives of Canadian youth. The repercussions for the victim's mental and physical well-being manifest in problems ranging from alienation, identity issues, deterioration of psychological and physical health to cyber and in-person bullying, and much more. The study concludes that while there are definite links between the rise of online hate speech, deterioration of mental and physical health, and increased attacks on immigrants and minorities, not much action has gone into policymaking and education to correct the situation.

Effects of tonotopic matching and spatial cues on segregation of competing speech in simulations of bilateral cochlear implants.

In the clinical fitting of cochlear implants (CIs), the lowest input acoustic frequency is typically much lower than the characteristic frequency associated with the most apical electrode position, due to the limited electrode insertion depth. For bilateral CI users, electrode positions may differ across ears. However, the same acoustic-to-electrode frequency allocation table (FAT) is typically assigned to both ears. As such, bilateral CI users may experience both intra-aural frequency mismatch within each ear and inter-aural mismatch across ears. This inter-aural mismatch may limit the ability of bilateral CI users to take advantage of spatial cues when attempting to segregate competing speech. Adjusting the FAT to tonotopically match the electrode position in each ear (i.e., increasing the low acoustic input frequency) is theorized to reduce this inter-aural mismatch. Unfortunately, this approach may also introduce the loss of acoustic information below the modified input acoustic frequency. The present study explored the trade-off between reduced inter-aural frequency mismatch and low-frequency information loss for segregation of competing speech. Normal-hearing participants were tested while listening to acoustic simulations of bilateral CIs. Speech reception thresholds (SRTs) were measured for target sentences produced by a male talker in the presence of two different male talkers. Masker speech was either co-located with or spatially separated from the target speech. The bilateral CI simulations were produced by 16-channel sinewave vocoders; the simulated insertion depth was fixed in one ear and varied in the other ear, resulting in an inter-aural mismatch of 0, 2, or 6 mm in terms of cochlear place. Two FAT conditions were compared: 1) clinical (200–8000 Hz in both ears), or 2) matched to the simulated insertion depth in each ear. Results showed that SRTs were significantly lower with the matched than with the clinical FAT, regardless of the insertion depth or spatial configuration of the masker speech. The largest improvement in SRTs with the matched FAT was observed when the inter-aural mismatch was largest (6 mm). These results suggest that minimizing inter-aural mismatch with tonotopically matched FATs may benefit bilateral CI users’ ability to segregate competing speech despite substantial low-frequency information loss in ears with shallow insertion depths.