Long-term Speech-shaped Noise Research Articles

Perceptual effects of formant enhancement with the factors of phonetic type, listening conditions, and language experience of listeners

The second formant (F2) enhancement is a technique that aims to improve speech perception in adverse noise by amplifying the F2 of speech signals. The current study was to investigate whether F2 enhancement would improve speech identification with the factors of phonetic type (e.g., vowel and consonant), listening conditions (e.g., speech and nonspeech noise at moderately and challenging SNRs), and language experience of listeners (e.g., native and nonnative listeners), and whether the amount of perceptual benefit was dependent on these factors. Two groups of participants, English native and nonnative listeners, were recruited in this study. Identification of English vowels and consonants with and without F2 enhancement were measured in quiet, long-term speech shaped noise (LTSSN) and six-talker babble (6-TB) at the signal-to-noise ratios (SNRs) of −10 dB and −15 dB. Overall, significant improvements from F2 enhancement were found in both vowel and consonant identification for both native and nonnative listeners in various listening conditions. Furthermore, greater improvement was found at the SNR of -15 dB than at the SNR of −10 dB, as well as for nonnative listeners than native listeners in vowel identification. Meanwhile, the amount of benefit was generally comparable speech and nonspeech noise. These results indicate that F2 enhancement could improve phonetic identification in noise for native and nonnative listeners, showing a potential as a speech enhancement algorithm in challenging noise.

Perceptual effects of formant enhancement with the factors of phonetic type, listening conditions, and language experience of listeners

The second formant (F2) enhancement is a technique that aims to improve speech perception in adverse noise by amplifying the F2 of speech signals. The current study was to investigate whether F2 enhancement would improve speech identification with the factors of phonetic type (e.g., vowel and consonant), listening conditions (e.g., speech and nonspeech noise at moderately and challenging SNRs), and language experience of listeners (e.g., native and nonnative listeners), whether the amount of perceptual benefit was dependent on these factors. Two groups of participants, English native and nonnative listeners, were recruited in this study. Identification of English vowels and consonants with and without F2 enhancement were measured in quiet, long-term speech-shaped noise (LTSSN) and six-talker babble (6-TB) at the signal-to-noise ratios (SNRs) of −10 dB and −15 dB. Overall, significant improvements from F2 enhancement were found in both vowel and consonant identification for both native and nonnative listeners in various listening conditions. Furthermore, greater improvement was found at the SNR of −15 dB than at the SNR of −10 dB, as well as for nonnative listeners than native listeners in vowel identification. Meanwhile, the amount of benefit was generally comparable in speech and nonspeech noise. These results indicate that F2 enhancement could improve phonetic identification in noise for native and nonnative listeners, showing a potential as a speech enhancement algorithm in challenging noise.

Re-evaluate the data of English vowel perception in quiet and noise for native and non-native listeners: Sensitivity and response bias

For decades, research on vowel perception has been primarily focused on the percentage correct of identification. Limited studies take listeners’ response bias into consideration and applied signal detection theory (SDT) in the data analysis of vowel identification. The goal of this study was to investigate English vowel identification in quiet, long-term speech-shaped noise (LTSS), and multi-talker babble (MTB) for English-native (EN), Chinese-native in the US (CNU), and Chinese-native in China (CNC) listeners by computing their sensitivity (measured by d’) and response bias (measured by c) using the SDT. Results showed that (1) in all three listening conditions, EN listeners showed higher sensitivity and lower bias than CNU and CNC listeners; (2) in quiet and MTB, CNU group demonstrated higher sensitivity than CNC group with similar sensitivity in LTSSN between two Chinese groups; and (3) in MTB, CNU group had smaller bias than CNC group, while in quiet and LTSSN condition, two Chinese groups showed similar bias. These results suggest that the US residency may not improve Chinese-native listeners’ vowel identification capacity, instead, may shift their response strategy toward native speakers’ pattern. [Work by the University of Texas at Austin Research Grant and China National Natural Science Foundation 31628009.]

Vowel discrimination in noise with formant enhancement: Effects of hearing loss and aging

The aim of this study was to investigate whether formant enhancement of F2 would improve vowel formant discrimination in noise for three groups of listeners (i.e., older listeners with hearing loss, older listeners with normal hearing, and young listeners with normal hearing). Thresholds of vowel formant discrimination were examined for the F2 frequencies of three American English vowels: /ʌ/, /ɛ/, and /i/ in long-term speech-shaped noise with signal-to-noise ratio of + 6 dB and + 12 dB. F2 frequencies of three vowels were enhanced by 6 dB. Thresholds of vowel formant discrimination were significantly highest for older listeners with hearing loss, while young listeners with normal hearing had lowest thresholds with the older listeners with normal hearing in between. However, discrimination thresholds were improved significantly for all three groups after F2 enhancement was implemented regardless of low or high SNR of background noise. Hearing loss and aging resulted in higher thresholds of vowel formant discrimination in noise and F2 enhancements could improve discrimination performance for all listeners in noise.

Psychometric Functions of Vowel Detection and Identification in Long-Term Speech-Shaped Noise.

Purpose The goal of this study was to investigate vowel detection and identification in noise and provide baseline data regarding how vowel perception changed with signal-to-noise ratios. Psychometric functions of vowel detection and identification for 12 American English isolated vowels in long-term speech-shaped noise were examined for young listeners with normal hearing in this study. Method Vowel detection was measured at sensation levels from -10 to +5 dB (re: thresholds of vowel detection from the study of Liu and Eddins, 2008a ) with a 4-interval forced-choice procedure. Thresholds of vowel detection were computed for each listener as the speech level at which 70.7% correct performance was reached. Vowel identification was then examined at sensation levels from 0 to 12 dB relative to detection thresholds for each listener. Thresholds of vowel identification were calculated as the speech level with vowel identifiability ( d') equals to 1. Results Thresholds of vowel detection and identification were significantly affected by vowel category. Slopes of psychometric functions of vowel identification were significantly dependent on vowel category, whereas slopes of psychometric functions of vowel detection were not. Conclusions These results suggest that, given the same sensation levels, especially at low sensation levels, vowel sounds are not equally perceivable in terms of identifiability.

Vowel perception in quiet and noise for native and non-native listeners: Analysis of vowel identifiability

For a number of years, research on phonetic perception has been primarily focused on the percentage correct of phonetic identification (e.g., hit rate) without considering listeners’ response (e.g., false alarm). The purpose of this study was to compute d’ of English vowel identification in quiet and noise including both hit rate and false alarm for native and non-native listeners. Both data analyses showed native listeners outperformed non-native listeners for vowel identification in quiet and noise. However, the d’ data showed that the non-naïve disadvantage became larger in long-term speech shaped (LTSS) noise than in quiet, while the percentage correct data showed an oppositely finding that the non-native disadvantage was smaller in LTSS noise than in quiet. These preliminary results suggest that false alarms should be included in the data analysis of phonetic perception and the idenitificability (e.g., d’) based on the signal detection theory may be more appropriate to evaluate listeners’ capacity to identify phonemes. [Work supported by China Natural Science Foundation 31628009 and University of Texas at Austin Research Grant].

English vowel identification in quiet and noise: effects of listeners' native language background.

Purpose: To investigate the effect of listener's native language (L1) and the types of noise on English vowel identification in noise.Method: Identification of 12 English vowels was measured in quiet and in long-term speech-shaped noise and multi-talker babble (MTB) noise for English- (EN), Chinese- (CN) and Korean-native (KN) listeners at various signal-to-noise ratios (SNRs).Results: Compared to non-native listeners, EN listeners performed significantly better in quiet and in noise. Vowel identification in long-term speech-shaped noise and in MTB noise was similar between CN and KN listeners. This is different from our previous study in which KN listeners performed better than CN listeners in English sentence recognition in MTB noise.Discussion: Results from the current study suggest that depending on speech materials, the effect of non-native listeners' L1 on speech perception in noise may be different. That is, in the perception of speech materials with little linguistic cues like isolated vowels, the characteristics of non-native listener's native language may not play a significant role. On the other hand, in the perception of running speech in which listeners need to use more linguistic cues (e.g., acoustic-phonetic, semantic, and prosodic cues), the non-native listener's native language background might result in a different masking effect.

English vowel identification in long-term speech-shaped noise and multi-talker babble for English and Chinese listeners

The identification of 12 English vowels was measured in quiet and in long-term speech-shaped noise (LTSSN) and multi-talker babble for English-native (EN) listeners and Chinese-native listeners in the U.S. (CNU) and China (CNC). The signal-to-noise ratio was manipulated from -15 to 0 dB. As expected, EN listeners performed significantly better in quiet and noisy conditions than CNU and CNC listeners. Vowel identification in LTSSN was similar between CNU and CNC listeners; however, performance in babble was significantly better for CNU listeners than for CNC listeners, indicating that exposing non-native listeners to native English may reduce informational masking of multi-talker babble.

English sentence recognition in speech-shaped noise and multi-talker babble for English-, Chinese-, and Korean-native listeners

This study aimed to investigate English sentence recognition in quiet and two types of maskers, multi-talker babble (MTB) and long-term speech-shaped noise (LTSSN), with varied signal-to-noise ratios, for English-, Chinese-, and Korean-native listeners. Results showed that first, sentence recognition for non-native listeners was affected more by background noise than that for native listeners; second, the masking effects of LTSSN were similar between Chinese and Korean listeners, but the masking effects of MTB were greater for Chinese than for Korean listeners, suggesting possible interaction effects between the non-native listener's native language and speech-like competing noise in sentence recognition.

Audibility of American English vowels produced by English-, Chinese-, and Korean-native speakers in long-term speech-shaped noise

The purpose of this study was to evaluate whether there were significant differences in audibility of American English vowels in noise produced by non-native and native speakers. Detection thresholds for 12 English vowels with equalized durations of 170 ms produced by 10 English-, Chinese- and Korean-native speakers were measured for young normal-hearing English-native listeners in the presence of speech-shaped noise presented at 70 dB SPL. Similar patterns of vowel detection thresholds as a function of the vowel category were found for native and non-native speakers, with the highest thresholds for /u/ and /ʊ/ and lowest thresholds for /i/ and /e/. In addition, vowel detection thresholds for non-native speakers were significantly lower and showed greater speaker variability than those for native speakers. Thresholds for vowel detection predicted from an excitation-pattern model corresponded well to behavioral thresholds, implying that vowel detection was primarily determined by the vowel spectrum regardless of speaker language background. Both behavioral and predicted thresholds showed that vowel audibility was similar or even better for non-native speakers than for native speakers, indicating that vowel audibility did not account for non-native speakers’ lower-than-native intelligibility in noise. Effects of non-native speakers’ English proficiency level on vowel audibility are discussed.

Sentence recognition in noise for English-, Chinese-, and Korean-native listeners.

The present study aims to investigate the effects of background noise on English sentence recognition for English-native (EN), Chinese-native (CN), and Korean-native (KN) listeners. The hearing in Noise Test (HINT) [Nilsson et al., (1994)] was used to measure the percent correct word identification in both quiet and noise conditions. Each sentence consists of uniform length with six syllables spoken by a female talker with general unaccented dialect-free American English. Two types of noise, multitalker babble and long-term speech shaped noise (LTSSN), were presented at various singal-to-noise ratios SNRs. Preliminary data demonstrated that first, the sentence recognition of non-native listeners was significantly lower than native listeners in both quiet and noise conditions; second, both native and non-native listeners showed significant amount of masking release at each SNR (−10, −5, and 0 dB SNRs); third, there was substantial individual variability in sentence recognition within each non-native group. This variability might be related with high variability in phoneme recognition in noise for the non-native individuals.

Psychometric properties of the coordinate response measure corpus with various types of background interference.

The coordinate response measure (CRM) corpus has gained broad acceptance as a research tool for investigating speech intelligibility in background competition and has been widely used in studies of informational masking. Beneficial properties of the CRM include closed set format, limited linguistic content, and inclusion of target items that promote sound stream segregation based on the voice-identify of the target talker. Extensive research has demonstrated the psychometric characteristics of CRM target-word identification in the presence of one, two, or three competing talkers from the same corpus as well as long-term speech-shaped noise (LTSSN). Percent correct identification varies non-monotonically with signal-to-background ratio for one and two CRM-talker interference and is monotonic when the number of talkers exceeds 2 or is LTSSN. The CRM has even broader applicability when used with other types of background competition. The purpose of this study is to establish the psychometric characteristics of CRM target-word identification in various backgrounds with the goal of being able to determine when it is appropriate or not to use adaptive threshold procedures with the CRM corpus. Psychometric functions and thresholds based on adaptive tracking for identification of CRM target words presented in multi-talker babble and cafeteria noise will be discussed.

Improvement of consonant identification in noise by spectro-temporal enhancement.

Our previous studies have shown that perception of vowels and consonants in long-term speech-shaped noise (LTSSN) by listeners with normal and impaired hearing is improved significantly by spectro-temporal enhancement resulting from modifications in the spectral modulation domain. The present study manipulated several parameters of enhancement processing including length of the analysis window and size of fast Fourier transform (FFT) analysis to determine which parameters produce the greatest improvement in stop consonant identification in the presence of LTSSN and cafeteria noise for normal-hearing listeners. The speech signals were presented at 70-dB sound pressure level and the signal-to-noise ratio was varied in fixed steps. Results indicated that identification of stop consonants increased by about 10 percentage points in LTSS noise and by about 25 percentage points in cafeteria noise when the length of analysis window was 200 ms and the FFT size was 512 points. Subsequent acoustic analyses showed that such perceptual improvement may be due to enhancement of formant transitions relative to background competition.

Vowel detection and vowel identification in long-term speech-shaped noise.

Psychometric functions of vowel detection and vowel identification were measured in long-term speech-shaped noise (LTSSN) for normal-hearing listeners. A four-interval forced-choice procedure was used to examine the accuracy of vowel detection in LTSSN with speech level presented from −10 to +5 dB sensational level relative to vowel detection thresholds obtained from Liu and Eddins’ study [J. Acoust. Soc. Am. 123, 4539–4546 (2008)]. The accuracy of vowel detection was significantly influenced by vowel category and sensational level. The threshold of vowel detection for each vowel and each listener was defined as the speech level at which 70.7% accuracy of vowel detection was reached. Vowel identification was then measured in LTSSN with vowel levels presented from 0 to 12 dB sensational level relative to individual thresholds of vowel detection, using a close-set 12-choice procedure. Results suggest that vowel identification was significantly affected by vowel category and sensational level. Altogether, the results of vowel detection and vowel identification indicate that, given the same signal-to-noise ratio, vowels are not equally audible and identifiable. Moreover, given the same sensational levels, vowels do not have the same identifiability, possibly due to the fact that some vowels have dominant confusing vowels while others do not.

Categorical dependence of vowel detection in long-term speech-shaped noise

The goal of this study was to measure detection thresholds for 12 isolated American English vowels naturally spoken by three male and three female talkers for young normal-hearing listeners in the presence of a long-term speech-shaped (LTSS) noise, which was presented at 70 dB sound pressure level. The vowel duration was equalized to 170 ms and the spectrum of the LTSS noise was identical to the long-term average spectrum of 12-talker babble. Given the same duration, detection thresholds for vowels differed by 19 dB across the 72 vowels. Thresholds for vowel detection showed a roughly U-shaped pattern as a function of the vowel category across talkers with lowest thresholds at /i/ and /ae/ vowels and highest thresholds at /u/ vowel in general. Both vowel category and talker had a significant effect on vowel detectability. Detection thresholds predicted from three excitation pattern metrics by using a simulation model were well matched with thresholds obtained from human listeners, suggesting that listeners could use a constant metric in the excitation pattern of the vowel to detect the signal in noise independent of the vowel category and talker. Application of the simulation model to predict thresholds of vowel detection in noise was also discussed.

Improvement of speech recognition thresholds by spectral modulation enhancement

Our previous studies have shown that vowel identification in noise was significantly improved by spectral enhancement resulting from modifications in the spectral modulation domain. The present study investigates whether spectral enhancement results in improved speech recognition in background competition. Speech recognition thresholds (SRT) were measured in a long-term speech shaped (LTSS) noise and in multi-talker babble using NU-6 words with and without spectral enhancement for young normal-hearing listeners. Results indicated that SRTs were significantly reduced, by 2.2 dB, in LTSS noise and in babble when speech sounds were spectrally enhanced between 2.0 and 2.5 cycles/octave, while no significant changes in SRTs occurred for enhancement between 1.5 and 2 cycles/octave or between 1.5 and 2.5 cycles/octave. Spectral modulation transfer functions were also measured for each listener. Linear regression analyses showed that SRT improvement was moderately correlated with modulation detection thresholds at a spectral modulation frequency of 0.5 cycles/octave. Thus, it appears that better modulation detection at low spectral modulation frequencies is associated with greater improvement in word recognition in background competition.

Enhancement of spectral modulation frequency in acoustics and identification of vowels

The pattern of intensity variations across audio frequency can be represented as a series of spectral modulation frequencies. Selectivity for spectral modulation frequency in the auditory system has been demonstrated by physiological and psychophysical studies, suggesting that auditory perception of complex sounds might be based on spectral envelope channels. The present study investigated the effects of enhancing spectral modulation frequency on vowel acoustics and vowel identification. Twelve naturally spoken American-English vowels by three female talkers served as the standard stimuli. Various spectral modulation frequency regions of the standard vowels combined with a long-term speech-shaped (LTSS) noise were enhanced. Identification performance for the vowels presented in LTSS noise with and without spectral modulation frequency enhancement was estimated for seven young normal-hearing listeners. Vowels were presented at a sensation level of 3 dB in the LTSS noise (70 dB SPL). Results indicated that vowel identification performance was improved on average by 4%–6% over the 12-vowel categories for spectral modulation enhancement between 1.0 and 2.5 cycles/octave. Acoustic analysis suggested that the improved vowel identification as a result of the global spectral modulation enhancement may be due to increases in the spectral contrast of individual formants.

Detection of vowels in long-term speech-shaped noise

The method of calibration is a critical issue in studies of vowel identification in noise, especially when manipulating speech-to-noise ratio. Common methods focus on stable estimates of the intensity or sound pressure level for specific speech sounds. However, vowels with the same sound pressure level often differ in both loudness, when presented at supra-threshold levels, and audibility, when masked detection thresholds for the vowel stimuli are measured. In the present study, detection thresholds for 12 isolated American-English vowels naturally spoken by three male and three female talkers were measured for young normal-hearing listeners in the presence of a long-term speech shaped (LTSS) noise. The duration of the vowels was equalized to 170 ms. The spectrum of the LTSS noise was identical to the long-term average spectrum of a 12-talker babble. The LTSS noise was present at 70 dB SPL. Thresholds for vowel detection showed distinct pattern as a function of the vowel category that was similar across listeners and talkers. This indicates that certain vowels were more audible than others, regardless of listeners and talkers. To account for the vowel-detection thresholds in noise, various analyses including comparisons of excitation patterns and estimates of specific loudness will be considered.

Formant discrimination in noise for isolated vowels

Formant discrimination for isolated vowels presented in noise was investigated for normal-hearing listeners. Discrimination thresholds for F1 and F2, for the seven American English vowels /i, I, epsilon, ae, [symbol see text], a, u/, were measured under two types of noise, long-term speech-shaped noise (LTSS) and multitalker babble, and also under quiet listening conditions. Signal-to-noise ratios (SNR) varied from -4 to +4 dB in steps of 2 dB. All three factors, formant frequency, signal-to-noise ratio, and noise type, had significant effects on vowel formant discrimination. Significant interactions among the three factors showed that threshold-frequency functions depended on SNR and noise type. The thresholds at the lowest levels of SNR were highly elevated by a factor of about 3 compared to those in quiet. The masking functions (threshold vs SNR) were well described by a negative exponential over F1 and F2 for both LTSS and babble noise. Speech-shaped noise was a slightly more effective masker than multitalker babble, presumably reflecting small benefits (1.5 dB) due to the temporal variation of the babble.

Vowel formant discrimination in quiet, speech-shaped noise, and babble for natural speech

Vowel formant discrimination in quiet and in noise for naturally spoken vowels was investigated. The stimuli were seven isolated vowels /i, u, I, ε, ʌ, æ, a/ recorded from a female talker and resynthesized using STRAIGHT with and without formant shifts. Vowels were presented in two noise types, long-term speech-shaped noise (LTSSN) and twelve-talker babble. ΔF was measured for F1 and F2 in quiet and noise with five signal to noise ratios (SNRs), +4, +2, 0, −2, and −4 dB. Adaptive tracking procedures were used, interleaving quiet and SNR levels for one formant at a time. Overall, results suggested that formant frequency, SNR, and noise type had significant effects on ΔF. For either LTSSN or babble, ΔF as a function of formant frequency showed four significantly different patterns, depending on the SNRs. ΔF as a function of SNR also showed several different patterns for different formants. A surprising result was that ΔFs were higher at −2 dB than those at −4 dB in babble, particularly in the F2 region. In addition, performance in babble was significantly better than that in LTSSN at comparable SNRs. These results will be examined by partial loudness models. [Work supported by NIHDCD-02229.]