增強語音理解度之結合時頻域增強之語音強化演算法

Combining Spectral and Temporal Speech Enhancement to Improve Speech Intelligibility

林世庭(Shih-Ting Lin)

指導教授 ： 冀泰石

國立交通大學/電機學院/電信工程系所/碩士(2014年)

https://doi.org/10.6842/NCTU.2014.00823

摘要

聽損者對語音時域變化的敏感度和頻率的解析度都較正常聽力的人為低，Moore團
隊提出個人化增強時域變化演算法以補償前項，但許多文獻說明頻率解析度變差會使聽損者於雜訊環境下的語音理解度大幅降低，因此，我們針對頻率解析度變差提出頻譜銳化和諧波增強兩演算法以期補償其聽損情形，然後我們將其分別結合Moore團隊所提出的個人化增強時域變化的演算法，提出時頻域皆補強的兩種方法。實驗設計為量測受試者於兩種雜訊環境下的中文單字語音理解度，並比較高訊雜比與低訊雜比的差別，將正常語音經過個人化聽損模型給聽力正常者聽，以評估此二套演算法在四種條件下的效能，最後並實測於聽力受損者，驗證對於中文單字語音理解度的實際幫助。由聽力正常者的實驗結果可說明結合個人化增強時域變化與頻譜銳化以及結合個人化增強時域變化與諧波增強的演算法分別相較於只考慮增強時域變化的演算法，對於中文語音理解度均有更進一步的提升；實測於聽損者的結果證實經結合個人化增強時域變化與頻譜銳化處理過後所測得的平均中文語音理解度高於只經增強時域變化處理所測得的平均中文語音理解度。未來展望需要進行聽損者測試以確認結合個人化增強時域變化與諧波增強演算法能提升聽損者的語音理解度。

關鍵字

聽力受損 ； 語音增強演算法 ； 語音理解度

並列摘要

The hearing-impaired have poorer sensitivities to temporal structures and frequency components than normal-hearing listeners. Moore's research group developed a personalized algorithm, which enhances spectral changes over time, to compensate the poor temporal resolution of the hearing-impaired. In addition, many studies have shown that the poor frequency resolution of the hearing impaired significantly degrades their speech perception in noise. In this thesis, we propose two spectral-temporal enhancement algorithms by combining the algorithm from Moore's group with the spectral-sharpening and harmonic-enhancing algorithms for the hearing-impaired. To validate our proposed enhancement algorithms, normal-hearing subjects are first asked to hear speech signals processed by proposed algorithms and the personalized hearing-impaired model under several noisy conditions. Then, the proposed algorithms are validated on actual hearing-impaired patients.

Experiment results of normal-hearing subjects demonstrate that our proposed algorithms, which combine the algorithm from Moore’s group with spectral-sharpening and harmonic-enhancing algorithms, produce more intelligible Mandarin speech than the algorithm from Moore’s group. Experiment results of hearing-impaired patients also demonstrate that the Mandarin intelligibility improvement of our proposed algorithm, the one which combines the algorithm from Moore’s group with the spectral-sharpening algorithm, is higher than the improvement of the algorithm from Moore’s group. The Mandarin intelligibility improvement of the other algorithm, which combines the algorithm from Moore’s group with the harmonic-enhancing algorithm, to the hearing-impaired patients will be assessed in the future.

並列關鍵字

hearing-impaired ； speech enhancement ； speech intelligibility

參考文獻

[1] P. J. Fitzgibbons, and F. L. Wightman, “Gap detection in normal and hearing-impaired listeners,” J. Acoust. Soc. Am., 72, pp. 761-765, 1982.

[2] R. S. Tyler, A. Q. Summerfield, E. J. Wood, and M. A. Fernandes, “Psychoacoustic and phonetic temporal processing in normal and hearing-impaired listeners,” J. Acoust. Soc. Am., 72, pp. 740-752, 1982.

[3] B. R. Glasberg, B. C. J. Moore, and S. P. Bacon, “Gap detection and masking in hearing-impaired and normal-hearing subjects,” J. Acoust. Soc. Am., 81, pp. 1546-1556., 1987.

[4] M. Florentine, S. Buus, B. Scharf, and E. Zwicker, “Frequency selectivity in normally-hearing and hearing-impaired observers,” J. Speech Hear. Res., 23, pp. 643-669, 1980.

[5] B. C. J. Moore, and B. R. Glasberg, “Comparisons of frequency selectivity in simultaneous and forward masking for subjects with unilateral cochlear impairments,” J. Acoust. Soc. Am., 80, pp. 93-107, 1986.