噪声对元音和辅音线索的影响分析

需积分: 7 201 浏览量更新于2024-07-19 收藏 1.15MB PDF 举报

"这篇论文主要探讨了噪声对元音和辅音语音特征的影响，由美国德克萨斯大学达拉斯分校的Gaurang Parikh和Philipos C. Loizou教授合作完成。研究评估了多说话者噪音和语音形状噪声在不同信噪比(S/N)条件下对元音和停顿辅音频谱的影响。" 在语音处理领域，语音增强是一项关键技术，其目标是提升在噪声环境中的语音可理解性和清晰度。这篇论文深入分析了噪声如何改变元音和停顿辅音的声学特性，以及这些变化如何影响听者的感知。首先，研究发现噪声对元音的影响主要体现在三个方面：a) 在三个频率带内，噪声与干净语音之间的频谱包络差异；b) 噪声中可靠的一阶谐波(F1)和二阶谐波(F2)信息的存在；c) 爆破音的频率和斜率变化。声学分析结果显示，尽管噪声存在，F1的检测仍比F2更可靠，而元音频谱的最大差异发生在中频带。这表明，在极端噪声环境中，听者可能依赖相对准确的F1频率信息和部分F2信息来识别元音。其次，对于停顿辅音的识别，即使在低至-5dB的信噪比下，其识别率仍然保持较高水平。这表明，尽管噪声对元音识别造成了一定挑战，但辅音的识别在一定程度上具有更强的鲁棒性，能够在噪声环境下保持较好的性能。此外，论文还可能涉及噪声对语音信号处理算法的优化，如自适应滤波器、谱减法、基于统计模型的方法等，以减少噪声对语音质量的影响，并提高语音在噪声环境下的可理解性。这些技术在现代通信、听力辅助设备、语音识别系统以及智能音频处理设备中有着广泛的应用。 "The influence of noise on vowel and consonant cues"这篇论文揭示了噪声环境对语音信号特征的复杂影响，为改进语音增强算法提供了重要的理论依据，有助于设计出更高效、适应性强的语音处理技术，以应对各种噪声环境的挑战。

interrelated aims. The ﬁrst aim is to quantify the effect of

noise by means of acoustic analysis of the vowel and conso-

nant spectra. A subset of the aforementioned factors, known

to be important for vowel and stop-consonant identiﬁcation,

will be assessed by comparing the acoustic parameters 共e.g.,

spectral tilt, etc.兲 estimated in quiet with those estimated in

noise. The acoustic parameter comparisons are meant to an-

swer several questions, including the following. 共1兲 How are

the vowel spectral envelopes 共critical-band spectra兲 affected?

共2兲 How are the two formant frequencies 共F1 and F2兲, known

to be major cues to vowel recognition, affected? 共3兲 How are

the spectral tilt and frequency of the burst spectra affected?

The above, and other, questions will be answered quantita-

tively by performing acoustic analysis of vowels and stop

consonants embedded in −5 to 10 dB noise.

The second aim of this paper is to assess the perceptual

effect of noise on vowel and stop-consonant identiﬁcation.

This will be done by performing correlation analysis be-

tween the acoustic parameter values and the vowel/stop-

consonant identiﬁcation scores. The second aim will be ad-

dressed in experiment 1. The results from the acoustic

analysis and experiment 1 taken together will provide valu-

able insights on the cues used by listeners to understand

speech in noise. Knowing how noise affects the spectrum of

speech is important for several reasons. For one, such knowl-

edge could help us design better noise reduction algorithms

that could potentially improve hearing-impaired listeners’

speech understanding in noise. Secondly, it could help us

better understand which speech features are perceptually ro-

bust in additive noise, and consequently which features lis-

teners attend to when identifying vowels or consonants in

noise.

II. ACOUSTIC ANALYSIS

A. Method

1. Speech material

The vowel material consisted of the vowels in the

words: “heed, hid, hayed, head, had, hod, hud, hood, hoed,

who’d, heard.” The stimuli were drawn from a large multi-

talker vowel set used by Hillenbrand et al. 共1995兲. A total of

66 vowel tokens were used for acoustic analysis: 33 vowels

produced by male speakers and 33 vowels produced by fe-

male speakers. There were 6 tokens for each of the 11 vow-

els, 3 produced by male speakers and 3 by female speakers.

A total of 20 different male speakers and 23 female speakers

produced the 66 vowel tokens. Each speaker produced only a

subset of the 11 vowels. The vowels were sampled at

16 kHz. Table I gives the steady-state F1 and F2 values of

the vowel stimuli used in this study. The F1 and F2 values

were sampled at the steady-state portion of the vowel and

averaged across all speakers. The steady-state F1 and F2

values 共Table I兲 of the vowel stimuli used in this study were

provided by Hillenbrand et al. 共1995兲.

Consonant material consisted of the stop consonants in

VCV context, where V = /i a, u/ and C= / bdgptk/.The

stimuli were drawn from recordings made by Shannon et al.

共1999兲. A total of 36 consonant tokens were used for acoustic

analysis: 18 consonants 共6 stops⫻ 3 vowel contexts兲 pro-

duced by a male speaker and 18 consonants produced by a

female speaker. The consonants were sampled at 44.1 kHz.

2. Noise

Two types of noise were used, multi-talker babble 共two

male and two female talkers兲 and speech-shaped noise. The

babble was taken from the AudiTEC CD 共St. Louis兲 and was

sampled at 16 kHz. The speech-shaped noise 共sampled at

20 kHz兲 was constructed by ﬁltering white noise through a

60-tap FIR ﬁlter with a frequency response that matched the

long-term spectrum of the 11 male and 11 female vowels.

Noise was ﬁrst up-sampled to the sampling frequency of the

vowel/consonant materials and then added to the vowels at

−5, 0, 5, and 10 dB. Figure 1 shows the averaged long-term

spectra of the multi-talker babble and speech-shaped noise.

3. Acoustic analysis of vowels

Prior to the acoustic analysis, the complete vowel data

set was manually segmented to 关h Vowel d兴. The starting and

ending times of the vocalic nuclei were measured by hand

TABLE I. Mean F1 and F2 frequencies 共in Hz兲 of the vowels used in this study.

Had Hod Head Hayed Heard Hid Heed Hoed Hood Hud Who’d

F1 Male 627 786 555 438 466 384 331 500 424 629 319

Female 666 883 693 492 518 486 428 538 494 809 435

F2 Male 1910 1341 1851 2196 1377 2039 2311 868 992 1146 938

Female 2370 1682 1991 2437 1604 2332 2767 998 1102 1391 1384

FIG. 1. 共Color online兲 The long-term spectra of the multi-talker babble and

continuous speech-shaped noise used in this study as maskers.

3876 J. Acoust. Soc. Am., Vol. 118, No. 6, December 2005 G. Parikh and P. C. Loizou: Effect of noise on perception

剩余14页未读，继续阅读

Sunrise_sxit

粉丝: 0
资源: 4

噪声对元音和辅音线索的影响分析

Affective and associative qualities in the learning styles of Chicanos and Anglos

元音空间的数学模型_A mathematical model of the vowel space

vowel

Morphological Normalization of Vowel Images for Articulatory Speech Recognition

vowel.py

VOWEL-COUNT-PROGRAM

Vowel Chrome Extension-crx插件

Unveiling the Secrets of MATLAB if Statements: Mastering the Art of Conditional Judgment

最新资源