Improve Speech Recognition Research Articles

Influence of Postponed Follow-Up after Cochlear Implant Activation during the COVID-19 Pandemic on Aided Sound Field Detection and Speech Recognition

Introduction: The objective of this study was to assess the influence of postponing the first post-activation follow-up due to the COVID-19 pandemic on the aided sound field detection thresholds and speech recognition of cochlear implant (CI) users. Methods: A retrospective review was performed at a tertiary referral center. Two groups of adult CI recipients were evaluated: (1) patients whose first post-activation follow-up was postponed due to COVID-19 closures (postponed group; n = 10) and (2) a control group that attended recommended post-activation follow-ups prior to the COVID-19 pandemic (control group; n = 18). For both groups, electric thresholds were estimated at initial activation based on comfort levels and were measured behaviorally at subsequent post-activation follow-ups. For the control group, behavioral thresholds were measured at the 1-month follow-up. For the postponed group, behavioral thresholds were not measured until 3 months post-activation since the 1-month follow-up was postponed. The aided pure-tone average (PTA) and word recognition results were compared between groups at the 3-month follow-up and at an interim visit 2–9 weeks later. Results: At the 3-month follow-up, the postponed group had significantly poorer word recognition (23 vs. 42%, p = 0.027) and aided PTA (42 vs. 37 dB HL, p = 0.041) than the control group. No significant differences were observed between 3-month data from the control group and interim data from the postponed group. Conclusions: The postponed follow-up after CI activation was associated with poorer outcomes, both in terms of speech recognition and aided audibility. However, these detrimental effects were reversed following provision of an individualized map, with behaviorally measured electric threshold and comfort levels. While adult CI recipients demonstrate an improvement in speech recognition with estimated electric thresholds, the present results suggest that behavioral mapping within the initial weeks of device use may support optimal outcomes.

The impact of cochlear implantation on speech and language outcomes in children with asymmetric sensorineural hearing loss

IntroductionCochlear implantation (CI) is considered in children with asymmetric sensorineural hearing loss (ASHL), or severe to profound sensorineural hearing loss in one ear, and better hearing contralaterally. The benefits of a CI in these children are not well established. This study investigated the impact of unilateral CI on speech and language outcomes in pediatric patients with ASHL. MethodsThe retrospective chart review identified pediatric patients who underwent cochlear implantation for ASHL at a tertiary single center institution between 2014 and 2017. Children who underwent cochlear implantation in the poorer ear and maintained a hearing aid in the better ear were included, while children with cognitive impairment or prior cochlear nerve deficiency and auditory neuropathy were excluded. Speech and language measures were evaluated prior to implantation and at regular six-month intervals postoperatively. Ear specific Audiometry was evaluated at every visit and varied developmentally appropriate speech recognition tasks, including questionnaires, closed-set, and open-set measures, accounted for advancing childhood milestones. Specific measures of speech articulation, expressive and receptive language, and vocabulary were assessed including: Goldman Fristoe Test of Articulation (GFTA-3), Clinical Evaluation of Language Fundamentals (CELF-P2 or CELF-5), Receptive One-Word Picture Vocabulary Test (ROWPVT) and Expressive One Word Picture Vocabulary Test (EOWPVT). The mean difference and standard deviation of speech and language outcomes at preoperative and postoperative evaluations were calculated. The change in outcomes were further assessed over each post-operative time point. ResultsTwenty-six children with asymmetric sensorineural hearing loss were included in the analysis, with 11 (42%) having bilateral enlarged vestibular aqueducts (EVA). At time of CI, unaided mean Pure Tone Average (PTA-4) for the better ear was 62Db HL, and worse ear was 92Db HL. There was an improvement in speech recognition in the implanted ear as expected. All preoperative speech and language evaluations were conducted within 7 months of cochlear implantation. Twenty-three patients had at least 1 postoperative re-evaluation. Fourteen children had preoperative and postoperative assessments with the same tool, with 13 patients (93%) showing improvement in at least 1 domain. For patients with repeated assessment within 12 months of surgery, we observed an average improvement in standard scores of: GFTA-3 (+11.2), CELF-P2 Core Language (+10.3), CELF-5 Core Language (+7.0), ROWPVT (+2.8) and EOWPVT (+4.5). Individual results are described. ConclusionImprovement in speech and language measures can be demonstrated in children undergoing unilateral cochlear implantation for asymmetric sensorineural hearing losses. These children, who are not typical CI candidates, can benefit from a CI in the poorer ear.

Use of Auditory Training and Its Influence on Early Cochlear Implant Outcomes in Adults.

Assess associations between postcochlear implant (CI) auditory training and early outcomes related to speech recognition and CI quality of life (CIQOL). Longitudinal, prospective cohort. Tertiary academic center. Seventy-two adults undergoing cochlear implantation for bilateral severe-to-profound hearing loss. Self-reported use of three categories of auditory training post-CI activation: (1) face-to-face training (e.g., speech pathologist), (2) passive home-based training (e.g., listening to audiobooks), and (3) computer-based training (e.g., interactive software). Change in Consonant-Nucleus-Consonant phoneme (CNCp), CNC word (CNCw), AzBio sentences in quiet, and CIQOL-35 Profile global and domain scores from pre-CI to 3-month post-CI. Of 72 patients, 52 (72.2%) used an auditory training resource. Of all patients, 18.4% used face-to-face training, 58.3% passive home-based training, and 33.3% computer-based training. At 3 months post-CI, use of any training was associated with greater improvement in speech recognition (d-range = 0.57-0.85) and global and domain-specific CIQOL scores, except entertainment (d-range = -0.33 to 0.77). Use of computer-based training demonstrated the greatest effect, with larger improvements in speech recognition (CNCp: d = 0.69[0.03,1.35]; CNCw: d = 0.80[0.14,1.46]; AzBio: d = 1.11[0.44,1.77]) and global and all domain-specific CIQOL scores (d-range = 0.05-1.35). Controlling for age, sex, household income, and use of multiple training resources, computer-based training remained the strongest positive predictor of speech recognition and CIQOL improvement, with significant associations with CNCp (ß = 33.07[1,43,64.719]), AzBio (ß = 33.03[5.71,60.35]), and CIQOL-global (ß = 10.92[1.15,20.70]) score improvement. Our findings provide preliminary evidence-based recommendations for use of specific auditory training resources for new adult CI recipients. Auditory training, especially self-directed computer software, resulted in improved speech recognition and CIQOL outcomes after 3 months and are widely available for CI users.

Analysis of multisource non-linguistic sound recognition among cochlear implant subjects

Naturalistic sounds carry rich information related to situational context or subject/system properties that contribute towards an acoustic awareness of the ambient environment. Sensorineural hearing loss reduces the functionality of the cochlea, auditory nerve, or central auditory pathways leading to degradation in auditory processing. Cochlear Implants (CIs) have been widely used to restore impaired auditory function and research advancements have focused on improving speech recognition and overall hearing-related quality of life. However, relatively few studies have investigated non-linguistic Sound Recognition (SR) among CI subjects. In this study, the recognition of sounds in CI users is assessed in a competing condition involving at least two non-linguistic sound sources. Furthermore, an end-to-end audio source separation neural network, SuDoRM-RF, with negative and permutation invariance training, is used for audio source recovery and comparatively assessed for potential identification improvement of non-linguistic sounds among CI users. Objective metrics such as classification accuracy, scale-invariant signal-to-distortion-ratio, and other audio quality related measures are used for assessment and subjective evaluation and compared against listener testing with CI subjects. The proposed study can model multi-source non-linguistic sound problems, the cocktail party effect and potentially provide an effective simulation for realistic listening test scenarios. [Study supported by NIH DC010494-01A.]

Circular microphone array beamforming to improve speech recognition for virtual agents

Under funding from an NIH NIDCD SBIR phase 1 and phase 2 grant, mh acoustics built a 16-element concentric circular microphone array using both analog and digital MEMS microphones. The circular microphone array design was trademarked as the EigenEar® microphone array and the unique beamforming processing is covered by U.S. patent 8903106. The presentation aims at showing that a circular third-order differential beamformer is capable of improving the recognition rate of popular voice recognition services used by a number of commercially available virtual assistants. These services offer an API to a client allowing speech data to be uploaded, analyzed, and returned in text form to the client. Experiments were done to compare the recognition accuracy from two separate audio streams. The EigenEar microphone array has two output channels and one channel contained the output of one omnidirectional microphone in the array and the other channel contained the output of the beamformer. The two audio streams are time synchronous so both streams presented to the virtual assistant are time aligned. The talk will describe the EigenEar circular microphone array and how the experiment was setup and the ASR results obtained from the virtual agents.

Neural correlates of bilateral summation between natural and vocoded speech

The normal auditory system integrates redundant information across ears, which aids listeners to detect weak sounds. For patients with single sided deafness (SSD), cochlear implants (CI) have been attempted expecting improved speech recognition in challenging listening scenarios. However, it is not yet known whether CI recovers the bilateral summation for weak sound detection in SSD. The binaural benefit of soft-speech recognition was tested in six SSD CI patients and thirty normal hearing (NH) listeners. The task was identifying quietly presented (i.e., RMS ∼6 dB SL) monosyllabic words. CI patients were tested in three conditions: acoustic-ear only (A1), electric-ear only (E1), or both ears (A+E). NH participants were tested in four conditions: monaural natural speech (A1), monaural vocoded speech (E1), dichotic natural and vocoded speech (A+E), or dichotic natural speech (A2). Electroencephalography was recorded during the task. Bilateral conditions (A+E and A2), even when one ear was vocoded, resulted better accuracy than monaural conditions (A1 and E1). However, A+E condition exhibited smaller early auditory cortical responses (i.e., N1P2 complex) than any natural-speech conditions (A2 and even A1). This result may imply that bilateral summation benefits in SSD-CI users are achieved by later cognitive processing rather than immediate early integration.

Comparison of sound perception using CIS and ACE sound coding strategies in cochlear implants

Objectives to study the effect of ACE and CIS sound coding strategies on sound perception in patients with the cochlear implants system produced by Cochlear Limited.
 Material and methods. The study included 50 patients taking the rehabilitation course in the Astrakhan branch of the National Medical Research Center for Otorhinolaryngology of the Federal Medico-Biological Agency over the past 5 years (from 2014 to 2019). The group of subjects included children over 7 years old and adults, whose success in rehabilitation made it possible to perform a full range of tests. The patients underwent tonal threshold audiometry and speech audiometry in a free sound field; the results obtained were registered in special MS Excel tables and further analysed using statistical methods.
 Results. There were no statistically significant differences in hearing thresholds on tonal audiometry when using the coding strategies ACE and CIS, however, differences in speech perception were observed on average by 4.2%. The patients experienced in using hearing aids reported improved speech recognition, with scores varying within 5%.
 Conclusion. Using a higher-resolution coding strategy can significantly improve speech recognition, while lower-resolution coding is beneficial for patients with digital hearing aid experience.

Research on the application of intelligent speech recognition technology in medical big data fog computing system

ABSTRACT Through the establishment of a voice database and voice recognition model, combined with voice recognition processing methods, an intelligent voice recognition input system with medical big data and fog computing as the scene is developed to realise the application of intelligent voice recognition technology in the system, which can reduce the medical staffdaily burden, reduce repetitive workload, and improve the quality and efficiency of clinical work. Based on the key technologies of intelligent voice recognition and medical big data, combined with the characteristics of fog computing, the development of electronic medical records and examination reports intelligent voice entry, mobile nursing intelligent voice entry, non-contact intelligent voice data interaction system, through the construction of fog computing model, design and develop an intelligent medical system to realise the rapid input and output of medical records and examination reports, reduce the workload of doctors and nurses, and concentrate their limited time on the treatment process itself. At the same time, according to the working characteristics of different medical departments, the intelligent system design adopts convenient, practical and efficient intelligent speech recognition technology application tools, and makes full use of big data technology and fog computing architecture to improve speech recognition rate, reduce network delay, and improve query efficiency, It is beneficial to assist the work of various clinical departments, better improve the treatment effect and innovate new medical models.

Effect of Increased Daily Cochlear Implant Use on Auditory Perception in Adults.

Purpose Despite the recommendation for cochlear implant (CI) processor use during all waking hours, variability in average daily wear time remains high. Previous work has shown that objective wear time is significantly correlated with speech recognition outcomes. We aimed to investigate the causal link between daily wear time and speech recognition outcomes and assess one potential underlying mechanism, spectral processing, driving the causal link. We hypothesized that increased CI use would result in improved speech recognition via improved spectral processing. Method Twenty adult CI recipients completed two study visits. The baseline visit included auditory perception testing (speech recognition and spectral processing measures), questionnaire administration, and documentation of data logging from the CI software. Participants watched an educational video, and they were informed of the compensation schedule. Participants were then asked to increase their daily CI use over a 4-week period during everyday life. Baseline measures were reassessed following the 4-week period. Results Seventeen out of 20 participants increased their daily CI use. On average, participants' speech recognition improved by 3.0, 2.4, and 7.0 percentage points per hour of increased average daily CI use for consonant-nucleus-consonant words, AzBio sentences, and AzBio sentences in noise, respectively. Questionnaire scores were similar between visits. Spectral processing showed significant improvement and accounted for a small amount of variance in the change in speech recognition values. Conclusions Improved consistency of processor use over a 4-week period yielded significant improvements in speech recognition scores. Though a significant factor, spectral processing is likely not the only mechanism driving improvement in speech recognition; further research is warranted.

Cognitive and Physiological Measures of Listening Effort During Degraded Speech Perception: Relating Dual-Task and Pupillometry Paradigms.

Purpose Listening effort is quickly becoming an important metric for assessing speech perception in less-than-ideal situations. However, the relationship between the construct of listening effort and the measures used to assess it remains unclear. We compared two measures of listening effort: a cognitive dual task and a physiological pupillometry task. We sought to investigate the relationship between these measures of effort and whether engaging effort impacts speech accuracy. Method In Experiment 1, 30 participants completed a dual task and a pupillometry task that were carefully matched in stimuli and design. The dual task consisted of a spoken word recognition task and a visual match-to-sample task. In the pupillometry task, pupil size was monitored while participants completed a spoken word recognition task. Both tasks presented words at three levels of listening difficulty (unmodified, eight-channel vocoding, and four-channel vocoding) and provided response feedback on every trial. We refined the pupillometry task in Experiment 2 (n = 31); crucially, participants no longer received response feedback. Finally, we ran a new group of subjects on both tasks in Experiment 3 (n = 30). Results In Experiment 1, accuracy in the visual task decreased with increased signal degradation in the dual task, but pupil size was sensitive to accuracy and not vocoding condition. After removing feedback in Experiment 2, changes in pupil size were predicted by listening condition, suggesting the task was now sensitive to engaged effort. Both tasks were sensitive to listening difficulty in Experiment 3, but there was no relationship between the tasks and neither task predicted speech accuracy. Conclusions Consistent with previous work, we found little evidence for a relationship between different measures of listening effort. We also found no evidence that effort predicts speech accuracy, suggesting that engaging more effort does not lead to improved speech recognition. Cognitive and physiological measures of listening effort are likely sensitive to different aspects of the construct of listening effort. Supplemental Material https://doi.org/10.23641/asha.16455900.

Restoration of High Frequency Auditory Perception After Robot-Assisted or Manual Cochlear Implantation in Profoundly Deaf Adults Improves Speech Recognition.

Background and Purpose: Robot-assisted cochlear implantation has recently been implemented in clinical practice; however, its effect on hearing outcomes is unknown. The aim of this preliminary study was to evaluate hearing performance 1 year post-implantation whether the electrode array was inserted manually or assisted by a robot.Methods: Forty-two profoundly deaf adults were implanted either manually (n = 21) or assisted by a robot (RobOtol®, Collin, Bagneux, France) with three different electrode array types. Participants were paired by age, and electrode array type. The scalar position of the electrode array in the cochlea was assessed by 3D reconstruction from the pre- and post-implantation computed tomography. Pure-tone audiometry and speech perception in silence (percentage of disyllabic words at 60 dB) were tested on the implanted ear 1 year post-implantation in free-field conditions. The pure-tone average was calculated at 250–500–750 Hz, 500–1,000–2,000–3,000 Hz, and 3,000–4,000–8,000 Hz for low, mid, and high frequencies, respectively.Results: One year after cochlear implantation, restoration of the high-frequency thresholds was associated with better speech perception in silence, but not with low or mid frequencies (p < 0.0001; Adjusted R2 = 0.64, polynomial non-linear regression). Although array translocation was similar using either technique, the number of translocated electrodes was lower when the electrode arrays had been inserted with the assistance of the robot compared with manual insertion (p = 0.018; Fisher's exact test).Conclusion: The restoration of high-frequency thresholds (3,000–4,000–8,000 Hz) by cochlear implantation was associated with good speech perception in silence. The numbers of translocated electrodes were reduced after a robot-assisted insertion.

Smartphone-based single-channel speech enhancement application for hearing aids.

This work presents a single-channel speech enhancement (SE) framework based on the super-Gaussian extension of the joint maximum a posteriori (SGJMAP) estimation rule. The developed SE algorithm is an open-source research smartphone-based application for hearing improvement studies. In this algorithm, the SGJMAP-based estimation for noisy speech mixture is smoothed along the frequency axis by a Mel filter-bank, resulting in a Mel-warped frequency-domain SGJMAP estimation. The impulse response of this Mel-warped estimation is obtained by applying a Mel-warped inverse discrete cosine transform (Mel-IDCT). This helps in filtering out the background noise and enhancing the speech signal. The proposed application is implemented on an iPhone (Apple, Cupertino, CA) to operate in real time and tested with normal-hearing (NH) and hearing-impaired (HI) listeners with different types of hearing aids through wireless connectivity. The objective speech quality and intelligibility test results are used to compare the performance of the proposed algorithm to existing conventional single-channel SE methods. Additionally, test results from NH and HI listeners show substantial improvement in speech recognition with the developed method in simulated real-world noisy conditions at different signal-to-noise ratio levels.

Increased maximum power output may improve speech recognition with bone conduction hearing devices

Objective To investigate the influence of maximum power output of bone conduction hearing devices on speech recognition in quiet and in noise in experienced users of bone conduction hearing devices. Design Prospective, randomised cross-over investigation comparing speech recognition performance, subjective sound quality, and device preference between two bone conduction hearing devices with different maximum power outputs. Study sample Sixteen adult subjects with conductive or mixed hearing loss. Results Both speech recognition in quiet and speech recognition in noise improved significantly when using the device with high vs. lower maximum power output. Mean improvement in word recognition score in quiet was 10.5% and the mean speech reception threshold in noise improved by 0.9 dB SNR. Compared to the device with lower maximum power output, the sound quality was rated significantly higher with the device with high maximum power output, which was also the device of preference for 81% of the subjects. Conclusion Bone conduction hearing devices with higher maximum power output have the potential to improve speech recognition in both quiet and noisy listening environments.

Feature compensation based on the normalization of vocal tract length for the improvement of emotion-affected speech recognition

The performance of speech recognition systems trained with neutral utterances degrades significantly when these systems are tested with emotional speech. Since everybody can speak emotionally in the real-world environment, it is necessary to take account of the emotional states of speech in the performance of the automatic speech recognition system. Limited works have been performed in the field of emotion-affected speech recognition and so far, most of the researches have focused on the classification of speech emotions. In this paper, the vocal tract length normalization method is employed to enhance the robustness of the emotion-affected speech recognition system. For this purpose, two structures of the speech recognition system based on hybrids of hidden Markov model with Gaussian mixture model and deep neural network are used. To achieve this goal, frequency warping is applied to the filterbank and/or discrete-cosine transform domain(s) in the feature extraction process of the automatic speech recognition system. The warping process is conducted in a way to normalize the emotional feature components and make them close to their corresponding neutral feature components. The performance of the proposed system is evaluated in neutrally trained/emotionally tested conditions for different speech features and emotional states (i.e., Anger, Disgust, Fear, Happy, and Sad). In this system, frequency warping is employed for different acoustical features. The constructed emotion-affected speech recognition system is based on the Kaldi automatic speech recognition with the Persian emotional speech database and the crowd-sourced emotional multi-modal actors dataset as the input corpora. The experimental simulations reveal that, in general, the warped emotional features result in better performance of the emotion-affected speech recognition system as compared with their unwarped counterparts. Also, it can be seen that the performance of the speech recognition using the deep neural network-hidden Markov model outperforms the system employing the hybrid with the Gaussian mixture model.

Speech Recognition With Informational and Energetic Maskers in Patients With Single-Sided Deafness After Cochlear Implantation.

Purpose The aim of the study was to assess the susceptibility to energetic and informational masking in patients with single-sided deafness (SSD) with one normal-hearing (NH) ear and a cochlear implant (CI) in the contralateral ear, understand the effect on speech recognition when spatially separating noise and speech maskers, and investigate the influence of the CI in situations with energetic and informational masking. Method Speech recognition was measured in the presence of either a modulated speech-shaped noise or one of two competing speech maskers in 11 SSD-CI listeners. The speech maskers were manipulated with respect to fundamental frequency to consider the effect of different voices. Measurements were conducted in the unaided (NH) and aided (NHCI) conditions. Spatial release from masking (SRM) was calculated for each masker type and both listening conditions (NH and NHCI) by subtracting scores of the colocated target and masker condition (S0N0) from the spatially separated target and masker conditions (S0N≠0). Results Speech recognition was highly variable depending on the type of masker. SRM occurred in the unaided (NH) and aided (NHCI) conditions when the speech masker had the same gender as the target talker. Adding the CI improved speech recognition when this speech masker was ipsilateral to the NH ear. Conclusions The amount of informational masking is substantial in SSD-CI listeners with both colocated and spatially separated target and masker signals. The contribution of SRM to better speech recognition largely depends on the masker and is considerable when no differences in voices between the target and the competing talker occur. There is only a slight improvement in speech recognition by adding the CI.

Common Brain Substrates Underlying Auditory Speech Priming and Perceived Spatial Separation.

Under a “cocktail party” environment, listeners can utilize prior knowledge of the content and voice of the target speech [i.e., auditory speech priming (ASP)] and perceived spatial separation to improve recognition of the target speech among masking speech. Previous studies suggest that these two unmasking cues are not processed independently. However, it is unclear whether the unmasking effects of these two cues are supported by common neural bases. In the current study, we aimed to first confirm that ASP and perceived spatial separation contribute to the improvement of speech recognition interactively in a multitalker condition and further investigate whether there exist intersectant brain substrates underlying both unmasking effects, by introducing these two unmasking cues in a unified paradigm and using functional magnetic resonance imaging. The results showed that neural activations by the unmasking effects of ASP and perceived separation partly overlapped in brain areas: the left pars triangularis (TriIFG) and orbitalis of the inferior frontal gyrus, left inferior parietal lobule, left supramarginal gyrus, and bilateral putamen, all of which are involved in the sensorimotor integration and the speech production. The activations of the left TriIFG were correlated with behavioral improvements caused by ASP and perceived separation. Meanwhile, ASP and perceived separation also enhanced the functional connectivity between the left IFG and brain areas related to the suppression of distractive speech signals: the anterior cingulate cortex and the left middle frontal gyrus, respectively. Therefore, these findings suggest that the motor representation of speech is important for both the unmasking effects of ASP and perceived separation and highlight the critical role of the left IFG in these unmasking effects in “cocktail party” environments.

Meta-Analysis of Speech Recognition Outcomes in Younger and Older Adults With Cochlear Implants.

Purpose Meta-analyses were conducted to compare pre- to postoperative speech recognition improvements and postoperative scores after cochlear implantation in younger (< 60 years) and older (> 60 years) adults. Method Studies were identified with electronic databases and through manual search of the literature. In the primary analyses, effect sizes between pre- and postoperative scores for each age group were calculated using a formula appropriate for repeated-measures designs. Using the effect sizes, two separate meta-analyses using a random-effects restricted maximum likelihood model were conducted for experiments using word and sentence recognition stimuli in quiet. Secondary meta-analyses were conducted to examine average postimplant, percent correct word recognition, sentence recognition, and speech recognition in noise in studies that included both older and younger age groups. Traditional Hedges's g effect sizes were calculated between the two groups. Results For the primary analyses, experiments using word and sentence recognition stimuli yielded significant, large effect sizes for the younger and older adult cochlear implant recipients with no significant differences between the older and younger age groups. However, the secondary meta-analyses of postoperative scores suggested significant differences between age groups for stimuli in quiet and noise. Conclusions Although older and younger adults with implants achieve the same magnitude of pre- to postimplant speech recognition benefit in quiet, the overall postoperative speech recognition outcomes in quiet and noise are superior in younger over older adults. Strategies to mitigate these group differences are critical for ensuring optimal outcomes in elderly individuals who are candidates for cochlear implants.

Audiologic outcomes and complications of active middle ear implantation in older adults

Background A few studies have reported the use of middle ear implants (MEIs) in older adults. Objectives To evaluate the audiologic outcomes and postoperative complications of MEIs in older adults. Methods This retrospective study reviewed audiologic data and medical records from a single referral centre. We identified 34 patients aged ≥65 years who underwent middle ear implantation using the Vibrant SoundbridgeTM device. Preoperative audiometric testing and postoperative aided audiometric testing were performed to evaluate the functional hearing gain at 1 year postoperatively. Patients were divided into 2 groups depending on whether they underwent explantation because of poor hearing benefit. Results Follow-up duration ranged from <1 to 5.3 years. The functional gain with MEIs significantly improved relative to the preoperative air conduction thresholds at 0.5, 1, 2, and 4 kHz. Eight patients underwent explantation and 7 lost their external audio processor devices. Those who removed their implants because of the poor hearing (group 1) showed significantly worse hearing thresholds at 1 kHz and speech discrimination scores than the others (group 2). Conclusions and significance MEIs for auditory rehabilitation can provide improved speech recognition and significant functional gains in older adults. Patients must be given appropriate preoperative explanations regarding the expected outcomes.

Análisis de la potencia de salida y del rendimiento auditivo de los dispositivos auditivos osteointegrados

Background and objectiveOsseointegrated auditory devices are hearing gadgets that use the bone conduction of sound to produce hearing improvement. The mechanisms and factors that contribute to this sound transmission have been widely studied, however, there are other aspects that remain unknown, for instance, the influence of the processor power output. The aim of this study was to know if there is any relationship between the power output created by the devices and the hearing improvement that they achieve. Materials and methodsForty-four patients were implanted with a percutaneous Baha® 5 model. Hearing thresholds in pure tone audiometry, free-field audiometry, and speech recognition (in quiet and in noise) were measured pre and postoperatively in each patient .The direct bone conduction thresholds and the power output values from the processors were also obtained. ResultsThe pure tone average threshold in free field was 39.29dB (SD 9.15), so that the mean gain was 29.18dB (SD 10.13) with the device. This involved an air-bone gap closure in 63.64% of patients. The pure tone average threshold in direct bone conduction was 27.6dB (SD 10.91), which was 8.4dB better than the pure tone average threshold via bone conduction. The mean gain in speech recognition was 39.15% (SD 23.98) at 40dB and 36.66% (SD 26.76) at 60dB. The mean gain in the signal-to-noise ratio was −5.9dB (SD 4.32). On the other hand, the mean power output values were 27.95dB μN (SD 6.51) in G40 and 26.22dB μN (SD 6.49) in G60. When analysing the relationship between bone conduction thresholds and G40 and G60 values, a correlation from the frequency of 1,000Hz was observed. However, no statistically significant association between power output, functional gain or speech recognition gain was found. ConclusionsThe osseointegrated auditory devices generate hearing improvement in tonal thresholds and speech recognition, even in noise. Most patients closed the air-bone gap with the device. There is a direct relationship between the bone conduction threshold and the power output values from the processor, but only in mid and high frequencies. However, the relationship between power output and gain in speech recognition is weaker. Further investigation of contributing factors is necessary.

Neural methods for effective, efficient, and exposure-aware information retrieval

Neural networks with deep architectures have demonstrated significant performance improvements in computer vision, speech recognition, and natural language processing. The challenges in information retrieval (IR), however, are different from these other application areas. A common form of IR involves ranking of documents---or short passages---in response to keyword-based queries. Effective IR systems must deal with query-document vocabulary mismatch problem, by modeling relationships between different query and document terms and how they indicate relevance. Models should also consider lexical matches when the query contains rare terms---such as a person's name or a product model number---not seen during training, and to avoid retrieving semantically related but irrelevant results. In many real-life IR tasks, the retrieval involves extremely large collections---such as the document index of a commercial Web search engine---containing billions of documents. Efficient IR methods should take advantage of specialized IR data structures, such as inverted index, to efficiently retrieve from large collections. Given an information need, the IR system also mediates how much exposure an information artifact receives by deciding whether it should be displayed, and where it should be positioned, among other results. Exposure-aware IR systems may optimize for additional objectives, besides relevance, such as parity of exposure for retrieved items and content publishers. In this thesis, we present novel neural architectures and methods motivated by the specific needs and challenges of IR tasks. We ground our contributions with a detailed survey of the growing body of neural IR literature [Mitra and Craswell, 2018]. Our key contribution towards improving the effectiveness of deep ranking models is developing the Duet principle [Mitra et al., 2017] which emphasizes the importance of incorporating evidence based on both patterns of exact term matches and similarities between learned latent representations of query and document. To efficiently retrieve from large collections, we develop a framework to incorporate query term independence [Mitra et al., 2019] into any arbitrary deep model that enables large-scale precomputation and the use of inverted index for fast retrieval. In the context of stochastic ranking, we further develop optimization strategies for exposure-based objectives [Diaz et al., 2020]. Finally, this dissertation also summarizes our contributions towards benchmarking neural IR models in the presence of large training datasets [Craswell et al., 2019] and explores the application of neural methods to other IR tasks, such as query auto-completion.