Sentence-length Utterances Research Articles

Perceptual Discrimination of Speaking Style Under Cochlear Implant Simulation.

Objectives:Real-life, adverse listening conditions involve a great deal of speech variability, including variability in speaking style. Depending on the speaking context, talkers may use a more casual, reduced speaking style or a more formal, careful speaking style. Attending to fine-grained acoustic-phonetic details characterizing different speaking styles facilitates the perception of the speaking style used by the talker. These acoustic-phonetic cues are poorly encoded in cochlear implants (CIs), potentially rendering the discrimination of speaking style difficult. As a first step to characterizing CI perception of real-life speech forms, the present study investigated the perception of different speaking styles in normal-hearing (NH) listeners with and without CI simulation.Design:The discrimination of three speaking styles (conversational reduced speech, speech from retold stories, and carefully read speech) was assessed using a speaking style discrimination task in two experiments. NH listeners classified sentence-length utterances, produced in one of the three styles, as either formal (careful) or informal (conversational). Utterances were presented with unmodified speaking rates in experiment 1 (31 NH, young adult Dutch speakers) and with modified speaking rates set to the average rate across all utterances in experiment 2 (28 NH, young adult Dutch speakers). In both experiments, acoustic noise-vocoder simulations of CIs were used to produce 12-channel (CI-12) and 4-channel (CI-4) vocoder simulation conditions, in addition to a no-simulation condition without CI simulation.Results:In both experiments 1 and 2, NH listeners were able to reliably discriminate the speaking styles without CI simulation. However, this ability was reduced under CI simulation. In experiment 1, participants showed poor discrimination of speaking styles under CI simulation. Listeners used speaking rate as a cue to make their judgements, even though it was not a reliable cue to speaking style in the study materials. In experiment 2, without differences in speaking rate among speaking styles, listeners showed better discrimination of speaking styles under CI simulation, using additional cues to complete the task.Conclusions:The findings from the present study demonstrate that perceiving differences in three speaking styles under CI simulation is a difficult task because some important cues to speaking style are not fully available in these conditions. While some cues like speaking rate are available, this information alone may not always be a reliable indicator of a particular speaking style. Some other reliable speaking styles cues, such as degraded acoustic-phonetic information and variability in speaking rate within an utterance, may be available but less salient. However, as in experiment 2, listeners’ perception of speaking styles may be modified if they are constrained or trained to use these additional cues, which were more reliable in the context of the present study. Taken together, these results suggest that dealing with speech variability in real-life listening conditions may be a challenge for CI users.

Perception of relative pitch of sentence-length utterances.

Differences in relative pitch have several linguistic uses, such as tone and dynamic intonational patterns. In addition, listeners are also sensitive to average pitch differences both within and across male and female speakers. While much research uses isolated syllables to investigate how well listeners can distinguish differences in pitch, this study focuses on listeners' ability to identify relative pitch differences across whole utterances. Listeners are presented with 28 items of the same sentence spoken by female politicians and are asked to rate how low or high the pitch of the utterance is on a 5-point scale. Results indicate that listeners use all steps of the scale to categorize the mean F0 of the stimuli (range: 130-250 Hz) and not a "high/medium/low" system or one in which only the extremes are categorized as high or low.

An Application for Screening Gradual-Onset Age-Related Hearing Loss

Objective: Less than 15% of adults in the USA over age 70 receive hearing screening; less than 20% of adults with hearing loss receive any form of treatment. Reasons vary, but affordability and accessibility are major barriers to intervention and treatment. This study provides data supporting a new adult hearing screening measure (NSRT) that is self-administered, easy to use and focused on difficulties experienced in everyday speech communication. Methods: The NSRT test materials are sentence-length utterances containing phonetic contrasts. The test requires respondents to determine whether sentences printed on a computer monitor are the same/different from sentences delivered as auditory stimuli through the computer sound card. The test is administered in quiet and +5 dB SNR background noise. Study participants were 120 adults aged 18 - 88 years. Results: Data obtained from the NSRT testing experience are used to construct a pseudo audiogram. When the predicted hearing thresholds were compared with conventional, clinical puretone measures, the sensitivity and specificity of the NSRT screening measure were 95% and 87%, respectively; diagnostic accuracy was 91%. Conclusions: The NSRT can identify individuals with hearing loss through a simple screening process grounded in standards set by the American Speech-Language-Hearing Association. The NSRT is suitable for administration in clinical and nonclinical settings.

Applications of the NTID speech recognition test (NSRT®)

Objective: This study evaluated the diagnostic capabilities of an adaptive speech recognition protocol (NSRT®) that can be self-administered in non-clinical venues by listeners using internet-based software. Design: All participants were given an audiological evaluation, including pure-tone testing, and responded to the NSRT administered in quiet and + 5 dB SNR listening conditions. The NSRT test materials are sentence-length utterances containing phonetic contrasts, primarily minimal pairs. Study sample: Subjects were 123 adults with normal hearing to moderately severe sensorineural hearing loss (mean age = 55 years, SD = 23). Results: Performance on the NSRT is strongly related to pure-tone thresholds. Linear regression analyses support the utility of the NSRT as a proxy for clinically-obtained hearing thresholds across the octave frequencies 0.5 to 8 kHz, primarily for individuals in the − 10 to 55 dB HL range. Other NSRT results are linked to analyses of phonetic errors and components of aural rehabilitation. Conclusions: Among its numerous results, the NSRT yields quantitative predictions of frequency-specific hearing thresholds, provides insight into the phonetic errors that affect speech understanding in adults who suffer from sensorineural hearing loss, primarily in the − 10 to 55 dB HL range, and has implications for the design of individualized auditory training programs.

Children’s Recognition of Spectrally Degraded Cartoon Voices

Although the spectrally degraded input provided by cochlear implants (CIs) is sufficient for speech perception in quiet, it poses problems for talker identification. The present study examined the ability of normally hearing (NH) children and child CI users to recognize cartoon voices while listening to spectrally degraded speech. In Experiment 1, 5- to 6-year-old NH children were required to identify familiar cartoon characters in a three-alternative, forced-choice task without feedback. Children heard sentence-length utterances at six levels of spectral degradation (noise-vocoded utterances with 4, 8, 12, 16, and 24 frequency bands and the original or unprocessed stimuli). In Experiment 2, child CI users 4 to 7 years of age and a control sample of 4- to 5-year-old NH children were required to identify the unprocessed stimuli from Experiment 1. NH children in Experiment 1 identified the voices significantly above chance levels, and they performed more accurately with increasing spectral information. Practice with stimuli that had greater spectral information facilitated performance on subsequent stimuli with lesser spectral information. In Experiment 2, child CI users successfully recognized the cartoon voices with slightly lower accuracy (0.90 proportion correct) than NH peers who listened to unprocessed utterances (0.97 proportion correct). The findings indicate that both NH children and child CI users can identify cartoon voices under conditions of severe spectral degradation. In such circumstances, children may rely on talker-specific phonetic detail to distinguish one talker from another.

Calibrating the detection of spontaneous speech: From sentences to noun phrases

Many studies have examined the differences between speech that is produced spontaneously as opposed to read from a prepared script. Most of these studies have focused on prosodic measures taken from clauses, sentences, or connected discourse. Furthermore, studies have shown that listeners are able to identify the context of production when presented with sentence-length utterances. The current study examined whether a listener can identify the context for utterances that are briefer than a sentence. A set of 20 talkers (10 male) produced spontaneous descriptions of maps that they then read aloud in a separate session at least one week later. Pairs of sentences that matched in fluency across both contexts were selected, and listeners judged which member of a pair was produced spontaneously. In separate blocks, listeners heard either full sentences, sentence beginnings, sentence endings, or two-word noun phrases excised from sentences. Overall, listeners could identify the spontaneously produced utterances,...

Perception of Japanese Temporally-cued Contrasts by American English Listeners

American listeners' perception of Japanese contrasts of vowel length (e.g., kiro vs. kiiro), consonant length (e.g., kite vs. kitte) and syllable number/length (e.g., k(j)oo vs. kijoo) was examined. Stimuli consisted of sentence-length utterances produced by a native Japanese talker; five minimal pairs of each contrast type were included. Questions were: a) can American listeners with no Japanese experience discriminate these contrasts, b) are there differences in relative difficulty of the three contrast types, and c) do instructions and other task variables affect performance accuracy? A categorial AXB discrimination task was used in two studies: Study I tested the three contrast types presented in separate blocks with detailed instructions about what to listen for; Study 2 tested the contrast types presented randomly intermixed with no specific instructions about the nature of the contrasts. Results indicated that naive American listeners could discriminate all three contrast types well above chance, and that there were no significant overall differences in relative difficulty across contrast types. Performance was significantly poorer in Study 2 where listeners' attention had not been directed to the nature of the contrasts. In both studies, there was significant and large variability in accuracy across different minimal pairs of each contrast type.

Effects of talker and token variability on perceptual learning of dialect categories.

Dialect classification is difficult for naive listeners, but perceptual learning tasks using sentence-length utterances have been shown to produce modest improvements in performance. The goal of the current study was to explore perceptual learning by naive listeners in a speeded dialect classification task with shorter, word-length utterances. In a series of experiments, participants were trained in a two-alternative forced-choice speeded dialect classification task (Cleveland vs. Cincinnati) with feedback and were then tested in the same task with novel talkers and novel words without feedback to assess learning. Variability in the stimulus materials in the training phase, including the number of talkers from each dialect and the number of different tokens produced by each talker, was manipulated across experiments to determine how variation in the input affected perceptual learning of dialect categories. The results revealed that training materials consisting of utterances produced by multiple different talkers from each dialect with multiple different tokens produced by each talker led to a significant improvement in dialect classification performance compared to a baseline condition without training. These findings suggest that dialect classification performance can improve with training on short, word-length utterances, but that robust dialect category learning requires high variability stimulus materials.

Effects of Talker Gender on Dialect Categorization

The identification of the gender of an unfamiliar talker is an easy and automatic process for naïve adult listeners. Sociolinguistic research has consistently revealed gender differences in the production of linguistic variables. Research on the perception of dialect variation, however, has been limited almost exclusively to male talkers. In the present study, naïve participants were asked to categorize unfamiliar talkers by dialect using sentence-length utterances under three presentation conditions: male talkers only, female talkers only, and a mixed gender condition. The results revealed no significant differences in categorization performance across the three presentation conditions. However, a clustering analysis of the listeners' categorization errors revealed significant effects of talker gender on the underlying perceptual similarity spaces. The present findings suggest that naïve listeners are sensitive to gender differences in speech production and are able to use those differences to reliably categorize unfamiliar male and female talkers by dialect.

Rapid adaptation to foreign-accented English

This study explored the perceptual benefits of brief exposure to non-native speech. Native English listeners were exposed to English sentences produced by non-native speakers. Perceptual processing speed was tracked by measuring reaction times to visual probe words following each sentence. Three experiments using Spanish- and Chinese-accented speech indicate that processing speed is initially slower for accented speech than for native speech but that this deficit diminishes within one minute of exposure. Control conditions rule out explanations for the adaptation effect based on practice with the task and general strategies for dealing with difficult speech. Further results suggest that adaptation can occur within as few as two to four sentence-length utterances. The findings emphasize the flexibility of human speech processing and require models of spoken word recognition that can rapidly accommodate significant acoustic-phonetic deviations from native language speech patterns.

Speaking mode variability in multimodal speech production

The speech acoustics and the phonetically relevant motion of the face during speech are determined by the time-varying behavior of the vocal tract. A benefit of this linkage is that we are able to estimate face motion from the spectral acoustics during speech production using simple neural networks. Thus far, however, the scope of reliable estimation has been limited to individual sentences; network training degrades sharply when multiple sentences are analyzed together. While there is a number of potential avenues for improving network generalization, this paper investigates the possibility that the experimental recording procedures introduce artificial boundary constraints between sentence length utterances. Specifically, the same sentence materials were recorded individually and as part of longer, paragraph length utterances. The scope of reliable network estimation was found to depend both on the length of the utterance (sentence versus paragraph) and, not surprisingly, on phonetic content: estimation of face motion from speech acoustics was reliable for larger sentence training sets when sentences were recorded in continuous paragraph readings; and greater phonetic diversity reduced reliability.

Talker-specific learning in speech perception.

The effects of perceptual learning of talker identity on the recognition of spoken words and sentences were investigated in three experiments. In each experiment, listeners were trained to learn a set of 10 talkers' voices and were then given an intelligibility test to assess the influence of learning the voices on the processing of the linguistic content of speech. In the first experiment, listeners learned voices from isolated words and were then tested with novel isolated words mixed in noise. The results showed that listeners who were given words produced by familiar talkers at test showed better identification performance than did listeners who were given words produced by unfamiliar talkers. In the second experiment, listeners learned novel voices from sentence-length utterances and were then presented with isolated words. The results showed that learning a talker's voice from sentences did not generalize well to identification of novel isolated words. In the third experiment, listeners learned voices from sentence-length utterances and were then given sentence-length utterances produced by familiar and unfamiliar talkers at test. We found that perceptual learning of novel voices from sentence-length utterances improved speech intelligibility for words in sentences. Generalization and transfer from voice learning to linguistic processing was found to be sensitive to the talker-specific information available during learning and test. These findings demonstrate that increased sensitivity to talker-specific information affects the perception of the linguistic properties of speech in isolated words and sentences.

Exploring declination utilizing a dynamic systems approach.

Recent reports suggest the value of examining various aspects of speech production from a nonlinear ‘‘dynamic systems’’ point of view [e.g., Kelso and Tuller, J. Physiol. 246, R928–R935 (1984)]. Such an approach searches for behaviors that reflect generic properties of self-organizing, dynamic systems, such as phase transitions, order or control parameters, and fractal dimensions [e.g., R. J. Baken, ASHA Fall 1990 and J. Voice (in press); M. T. Turvey, Am. Psychol. 45, 938–953 (1990)]. The attempt to examine fundamental frequency declination and intrautterance resets over phrase and sentence length utterances using this approach will be reported. Using reiterant syllable utterances, effects of task variations, such as speaking rate, intonation, and depth of initial inspiration, are examined in the context of what changes in behavior reveal about the coupling of component systems and possible ‘‘order parameters.’’ [Work supported by the Dept. of Psychol. and Chancellor’s Fellowship.]

On temporal alignment of sentences of natural and synthetic speech

One way to improve the quality of synthetic speech, and to learn about temporal aspects of speech recognition, is to study the problem of time aligning pairs of spoken sentences. For example, one could evaluate various sets of duration rules for synthesis by comparing the time alignments of speech sounds within synthetic sentences to those of naturally spoken sentences. In this manner, an improved set of sound duration rules could be obtained by applying some objective measure to the alignment scores. For speech recognition applications, one could obtain automatic labeling of continuous speech from a hand-marked prototype to obtain models and/or statistical data on sounds within sentences. A key question in the use of automatic alignment of sentence length utterances is whether the time warping methods, developed for isolated word recognition, could be extended to the problem of time aligning sentence length utterances (up to several seconds long). A second key question is the reliability and accuracy of such an alignment. In this paper we investigate these questions. It is shown that, with some simple modifications, the dynamic time warping procedures used for isolated word recognition apply almost as well to alignment of sentence length utterances. It is also shown that, on the average, the uncertainty in the location of significant events within the sentence is much smaller than the event durations although the largest errors are longer than some event durations. Hence, one must apply caution in using the time alignment contour for synthesis or recognition applications.

Speaker independent connected word recognition using a syntax-directed dynamic programming procedure

A method for speaker independent connected word recognition is described. Speaker independence is achieved by clustering isolated word utterances of a 100 speaker population. Connected word recognition is based on a syntax-directed dynamic programming algorithm which matches the isolated word templates to sentence length utterances. The method has been tested on an artificial task-oriented language based on a 127 word vocabulary. Four subjects, two men and two women, spoke a total of 209 sentences comprising 1750 words. At an average speaking rate of 171 words/min over dialed-up telephone lines, a correct word recognition rate of 97 percent was observed.