语音发音中元音清晰度的感知与声学研究外文文献

Perception and Acoustic Studies of Vowel Intelligibility in Dysphonic Speech Keiko Ishikawa Charles Nudelman Searam Park and Corey Ketring Champaign Illinois Summary Purpose This study aimed to 1 examine the effect of dysphonia and background noise on the identifi ability of vowels and 2 evaluate the relationship between the degree of vowel intelligibility and vowel related acoustic measurements Method A speech perception experiment was conducted with speech samples collected from 10 adult females with healthy voices and 10 adult females with dysphonia The speech material was 11 vowels of American English in h vowel d format Cafeteria noise was added to these samples at a signal to noise ratio of 6 dB A total of 10 adults with normal hearing participated in a speech perception experiment in which the vowels were presented with and without the noise F1 and F2 frequencies of the vowels were measured and their relationships with the vowel intelligibility were statistically evaluated Results The group level analysis showed that vowel intelligibility was lower in dysphonic speech than normal speech both in quiet and at signal to noise ratio of 6 dB The intelligibility was higher for the high vowels than the low vowels In general the vowel confusion pattern was similar between normal and dysphonic speech How ever data from a speaker with severe diplophonia showed a distinct confusion pattern Voice quality ratings did not signifi cantly correlate with the vowel intelligibility There was a signifi cant correlation between F2 and the vowel intelligibility in quiet A post hoc acoustic analysis revealed that vowels of the speakers with lower vowel intelligibility had lower acoustic energy above 1 kHz Conclusion Dysphonia negatively affects vowel intelligibility Low vowels were more vulnerable to the effect of dysphonia than high vowels Among different types of dysphonic voice qualities diplophonia appears to be par ticularly detrimental to vowel intelligibility F2 signifi cantly correlated with vowel intelligibility however this result requires a careful interpretation Given that the acoustic energy above 1 kHz resulted in better intelligibility in noise a treatment strategy that targets this frequency region may help improve intelligibility in noise Future studies should examine the generalizability of this fi nding to different age and gender groups and intelligibility as a whole Key Words Dysphonia Speech intelligibility Background noise Acoustic analysis INTRODUCTION Individuals with dysphonia are perceivedto be less intelligible especially in noisy places 1While the intelligibility problem has been well recognized clinically there is minimal research on how dysphonia affects intelligibility Understanding this relationship would help clinicians develop a therapy program that directly targets intelligibility enhancement The lack of intelligibility in dysphonic speech is also concerning in terms of vocal health When speakers cannot maintain intelligibility of their speech in noise they tend to speak louder in an attempt to improve their intelligibility known asthe Lombard effect 2This increase in loudness is accompanied by greater subglottic pressure and collision force applied to the vocal fold tissue which ultimately raises the risk of phonotrauma 3 If clinicians can develop a therapy program that maximizes the patient s speech intelligibility without raising loudness it would help prevent further phonotrauma and potentially expedite the patient s recovery Scientifi c evidence in this area is emerging Recently Evitts et al4conducted a speech perception experiment with samples from fi ve healthy speakers and fi ve speakers with dysphonia secondary to phonotrauma Their results showed that the listeners require more time to process dysphonic speech in comparison to normal speech and misunderstand more words in dysphonic speech Our recent studies also showed that dysphonic speech is less intelligible than normal speech especially in the presence of background noise 5Our previous work also indicated that the intelligibility defi cit cannot be attributed to reduced loudness alone Dysphonic speech was still less intelligible than normal speech after equalizing their intensity 5 This fi nding implies that the tem poral or spectral changes in dysphonic speech may contrib ute to the defi cit While fi ndings from these studies are valuable these studies examined sentence level intelligibil ity preventing us from determining what about the dys phonic signal contributed to the defi cit Defi ning the contributing factor would help us design a targeted therapy program for improving intelligibility Linguistically speech sounds are categorized into two broad phonemic classes vowels and consonants Past studies with healthy speakers have shown that both types of sounds are important for listeners to understand speech yet vowels Accepted for publication December 23 2019 From the Department of Speech and Hearing Science College of Applied Health Sciences University of Illinois at Urbana Champaign Champaign Illinois Address correspondence and reprint requests to Keiko Ishikawa Department of Speech and Hearing Science College of Applied Health Sciences University of Illi nois at Urbana Champaign 901 South 6th Street Champaign IL 61820 E mail ishikak illinois edu Journal of Voice Vol that is listeners identify vowels more accu rately when their articulatory position is at the extreme end of the vocal tract The primary acoustic cues for vowels are the fi rst two for mants F1 and F2 which are spectral regions amplifi ed by the vocal tract 9 11 12 The formant frequencies are infl uenced by the physical dimension of the vocal tract The relation ship between the formant frequencies and vocal tract shape has been described primarily based on tongue height and advancement 13F1 corresponds to tongue height with the higher F1 associated with low positioning of the tongue body while F2 grossly corresponds to the anteroposterior placement of the tongue with high F2 associated with ante rior tongue placement However underlying physiology for F2 seems to be more complex than originally theorized 14 Laryngeal pathologies that underlie dysphonia could infl u ence the acoustic characteristics of vowels Laryngeal pathol ogy frequently results in incomplete vocal fold closure 15 When the glottis is incompletely closed it allows the coupling of supraglottic and tracheal resonances The coupling affects formant frequencies A study by Fujimura and Lindqvist16 showed that opening the glottis raised F1 frequency More recently Lulich17used computational simulation to study the relationship between glottal opening and formant frequencies The study showed that widening the glottal width introduced the fi rst lower airway resonance to the supraglottic vocal tract resonance affecting the frequency region around F1 and F2 The low back vowels were particularly sensitive to this effect in comparison to other vowels Such change in formants may affect the resultant perception of the vowels The author suggested that this sensitivity could potentially be used to evaluate voice pathologies Furthermore the glottal opening alsoresults in greater transglottal airfl ow whichgenerates tur bulent like high frequency noise This noise may mask acous tic information associated with formants The listeners will be further challenged when they need to understand this type of speech in the presence of back ground noise For example road traffi c noise has strong energy under 1 kHz 18which is a frequency range that corre sponds to F1 for all vowels in American English When this type of noise masks F1 listeners must depend on F2 or other acoustic and linguistic cues to identify the vowels However the robustness of F2 may already be compro mised by the laryngeal pathology leading to greater diffi culty in understanding speech Dysphonia likely reduces a listener s ability to recognize vowels however there is little information on this topic in the current literature Limited evidence lies in studies of whispered speech which could be considered an extreme case of dysphonia One of the examples is a study by Kal lail 19which compared the intelligibility of normally pho nated and whispered vowels by adult males The authors found a 15 decrease in the intelligibility of whispered vow els compared to normally phonated vowels Furthermore whispered vowels had higher F1 than normally phonated vowels though its relationship with the intelligibility decrease was not studied F2 and F3 frequencies were less affected by whispering Their study presented the vowels without background noise and whispered voice is different from dysphonic voice Therefore understanding the nature of the intelligibility defi cit in dysphonic speech requires its own investigation This study has two general areas of focus 1 to character ize the effect of dysphonia on vowel intelligibility and 2 to identify acoustic correlates of the vowel recognition problem in dysphonic speech The fi rst part of the study examined the effect of dysphonia background noise vowel type and speaker on vowel intelligibility Research Question 1a We further examined which speakers were affected by noise Research Question 1b and which vowels are more vulnera ble to the effect of noise and dysphonia Research Question 1c We also qualitatively examined whether there is a particu lar confusion pattern between vowels intended by the speakers and identifi ed by the listeners Research Question 1d Lastly we explored the relationship between voice quality rating and intelligibility Research Question 1e The second part of the study evaluated the effect of dysphonia on F1 and F2 frequen cies Research Question 2a and the relationship between the degree of vowel intelligibility and the formant frequencies Research Question 2b METHODS Participants Speakers A total of 10 adult females with normal voices and 10 adult females with mild to severely dysphonic voices participated in the study All speakers were between 20 and 60 years of age These speakers were selected from a pool of 29 speakers with dysphonia and 30 speakers with normal voices None of the speakers reported a history of speech and language disorders The cause of dysphonia was struc tural in its origins such as vocal fold nodules polyps and ARTICLE IN PRESS 2Journal of Voice Vol however none of them required a break FIGURE 1 The interface of the stimuli presentation program ARTICLE IN PRESS Keiko Ishikawa et alPerception and Acoustic Studies of Vowel Intelligibility in Dysphonic Speech3 Acoustic analysis Acoustic analysis was conducted with automated speech analysis algorithm script implemented to run on PRAAT 22 For the analysis of speech samples experimenters visually confi rmed a steady portion of the vowel on a spectro gram The fi rst and second formants F1 and F2 were mea sured every 6 25 milliseconds in this portion Average F1 and F2 values were then calculated from the measurements The acoustic characteristics of the cafeteria noise were described with the long term average spectrum LTAS analysis The alpha ratio was then calculated to examine the difference in acoustic power in lower ie 1 kHz and higher frequency ranges Statistical analysis Several statistical tests were conducted as described below In general intelligibility was defi ned as the percentage of correctly identifi ed vowels The interrater reliability between the SLPs on their voice quality rating was examined with Cohen s linearly weighted kappa For the experiment two dif ferent listener groups were assigned to two different noise con ditions ie quiet vs SNR 6 Their potential difference in the listener s vowel recognition ability was examined by com paring their performance on the practice set The difference in the percentage of correctly identifi ed vowels between the lis tener groups was evaluated with the Wilcoxon rank sum test The interrater reliability among listeners for the experimental set was evaluated with the intraclass correlation coeffi cient ICC separately for each noise condition Research question 1a did dysphonia background vowel type and speaker affect vowel intelligibility In order to evaluate the effect of dysphonia and noise on overall vowel intelligibility and to identify outliers the per centage of correct responses for the vowels was calculated for dysphonic and normal groups An outlier was defi ned as any data point more than 1 5 interquartile ranges below the fi rst quartile or above the third quartile The effects of dys phonia noise vowel and speaker were examined using the multi way analysis of variance ANOVA test The depen dent variable was the number of correct responses averaged over listeners The following hypothesis was tested Dyspho nia background noise vowel type and speaker do not affect vowel intelligibility Research question 1b which speaker s intelligibility was affected by noise Generalized linear mixed models were built for each speaker to examine whether the percentage of correct responses dif fered between quiet and noisy conditions The listener was set as the random effect for the intercept and and noise was set as a fi xed effect for these models The following hypothe sis was tested Noise does not affect vowel intelligibility of a speaker Research question 1c which vowel s intelligibility was affected by dysphonia in a noisy listening environment The generalized linear mixed models were applied to each vowel in each noise condition to examine whether the per centage of corrected responses differed between normal and dysphonic groups The listener was set as a random effect and dysphonia was set as a fi xed effect for these models The following hypothesis was tested dysphonia does not affect vowel intelligibility in noise Research question 1d what is the confusion pattern in the vowel intelligibility Confusion matrices were generated to visualize the pattern of confusion between target vowels and listeners responses for both speaker groups in each noise condition The matri ces were generated for the speaker groups as well as for each speaker Research question 1e does voice quality rating predict the vowel intelligibility Spearman rank correlation tests were conducted to examine the relationship between voice quality rating in each percep tual category ie grade roughness breathiness asthenia and strain and vowel intelligibility in quiet and noise Bon ferroni correction was applied to adjust the critical value for the multiple comparisons The P value of less than 0 01 was considered signifi cant The following hypothesis was tested voice quality rating does not predict vowel intelligibility Research question 2a does dysphonia affect formant frequencies of the vowels Wilcoxon Rank Sum tests were conducted for each vowel to evaluate the difference in the formant values between normal and dysphonic groups The values of all the speakers in each group were pooled The following hypothesis was tested There is no statistically signifi cant difference in F1 and F2 fre quencies between the dysphonic and normal groups Research question 2b do formant frequencies predict vowel intelligibility The formant values for each vowel were averaged across the speakers and then the difference between the normal and dys phonic groups was calculated ie DF1 and DF2 The percent age of correct responses were averaged across speakers and listeners The difference between dysphonic and normal groups was then calculated for each vowel ie Dintelligibility Spearman rank correlation tests were conducted for each for mant and noise condition in order to evaluate the correlation between the difference in formant values and the difference in vowel intelligibility ie DF1 and Dintelligibility DF2 and Dintelligibility The following hypothesis was tested The dif ference in F1 and the difference in F2 do not predict the differ ence in vowel intelligibility ARTICLE IN PRESS 4Journal of Voice Vol k 0 242 P 0 197 respectively Results from the perception experiment Difference between the listener groups The mean score of the quiet group and SNR 6 group on the practice set were 72 SD 16 43 and 68 SD 26 8 respectively Wilcoxon rank sum test indi cated that there is no signifi cant difference between the quiet group and SNR 6 group w 14 P 0 83 Inter rater reliability across the listeners The ICC for the quiet condition was 0 84 with a 95 confi dence interval from 0 81 to 0 87 F 219 876 27 P 0 001 The ICC for SNR 6 dB was 0 62 with a 95 confi dence interval from 0 57 to 0 68 F 219 876 9 2 P 0 001 Research question 1a the effect of dysphonia noise vowel type and speaker on vowel intelligibility The average vowel intelligibility across speakers listeners and vowel types were 76 SE 2 for dysphonic group and 88 SE 1 for normal group in quiet and 63 SE 2 for dysphonic group and 77 SE 2 for normal group in noise Figure 2 The multi way ANOVA test indi cated that the effects of dysphonia noise vowel and speaker are signifi cant F 1 409 34 49 P 0 01 for dysphonia F 1 409 28 11 P 0 01 for noise F 10 409 12 95 P 0 01 for vowel F 18 409 6 43 P 0 01 for speaker Research question 1b the effect of the speaker on intelligibility in noise The generalized linear mixed models indicated that the intelligibility was lower in noisy condition for four speakers in the dysphonic group D4 5 6 9 and two speakers in the normal group N3 and10 P 0 01 for all of these cases Figure 3 Inspection of the data revealed that D6 s intelli gibility was below the fi rst quartile indicated that this speaker was an outlier Research question 1c the effect of vowel type on intelligibility For the quiet condition the models indicated that the vowel intelligibility was lower in the dysphonic group for the vowels had head and hud P 0 05 for all of these vowels For the noisy condition the models indicated that the intelligibility was lower in the dysphonic group for the vowels hawd head heard and hud and o hoed P 0 01 Results from the acousti