汉语音段反转言语的可懂度研究

实验研究了帧长对汉语音段反转言语可懂度的影响。实验结果表明,帧长在64 ms以下,汉语音段反转言语具有较高的可懂度;帧长在64~203 ms之间,可懂度随帧长的增加逐渐降低;帧长在203 ms以上,可懂度为0。在帧长8 ms时,汉语的声调失真导致可懂度下降。原始语音信号和音段反转言语的调制谱的分析表明,调制谱失真大小和可懂度密切相关。因此,用原始语音信号和音段反转言语的窄带包络间的归一化相关值可以衡量调制谱失真大小,基于语音的语言传输指数法计算的客观值和实验结果显著相关(r=0.876,p<0.01)。研究表明,语言可懂度与窄带包络有关,音段反转言语的可懂度和保留原始语音信号的窄带包络密切相关。      

小波包自适应阈值语音降噪新算法

为了克服低信噪比输入下,语音增强造成语音清音中的弱分量损失,造成重构信号包络失真的问题。论文提出了一种新的语音增强方法。该方法根据语音感知模型,采用不完全小波包分解拟合语音临界频带,并对语音按子带能量进行清浊音区分处理,在阈值计算上,提出了一种清浊音分离,基于子带信号能量的小波包自适应阈值算法。通过仿真实验,客观评测和听音测试表明,该算法在低信噪比输入时较传统算法,能够更加有效地减少重构信号包络失真,在不损伤语音清晰度和自然度的前提下,使输出信噪比明显提高。将该算法与能量谱减法结合,进行二次增强能进一步提高降噪输出的语音质量。     

基于频率调制信息的人工耳蜗语音处理算法研究

在传统人工耳蜗连续交叠采样(Continuous Interleaved Sampler,CIS)算法的基础上,提出一种基于精细结构(频率调制信息)的人工耳蜗语音处理算法,在不引入过高频率成分、保证工艺可实现性的前提下,使语音识别率大幅提高。听觉仿真实验的结果表明,与传统的基于时域包络的CIS算法相比,基于精细结构的CIS算法对于元音可懂度的改进可以达到28%;声调的识别率在各种噪声条件下提高20%以上;在一般噪声环境下,辅音和句子的可懂度也分别获得了22.9%和28.3%的改进。      

采用低维特征映射的耳语音向正常音转换

在将耳语音转换为正常音时,为了研究降维后语音特征对耳语音转换的影响,分别对耳语音和正常音谱包络进行自适应编码以提取耳语音和正常音的低维特征,然后使用BP网络建立耳语音和正常音低维谱包络特征之间的映射关系以及正常音基频和耳语音低维谱包络特征之间的关系。转换时,根据耳语音低维谱包络特征获得对应正常音的低维谱包络特征和基频,对低维谱包络特征进行解码后获得对应的正常音谱包络。实验结果表明,采用此方法转换后的语音与正常音之间的倒谱距离相比高斯混合模型方法下降了10%,转换后语音的自然度和可懂度都有所提高。      

基于能量损失率估计的传声器阵列语音增强

在延时相加波束形成和维纳滤波技术的基础上,提出了一种基于能量损失率估计的传声器阵列后滤波语音增强算法。该算法通过检测线性不等间距传声器阵列中各嵌套子阵在波束形成前后的能量变化来估计维纳滤波器的权系数,实现了语音增强的目标。在仿真数据集上的实验评估表明,相比原始语音,该算法增强后的语音在信噪比、对数谱距离和感知质量等指标上平均分别改善了17.1 dB,1.001和0.935,具有很好的应用前景。      

单通道语音增强算法对汉语语音可懂度影响的研究

考察了当前常用的几种单通道语音增强算法对汉语语音可懂度的影响。受不同类型噪音干扰的语音经过5种单通道语音增强算法的处理后,播放给具有正常听力水平的被试进行听辩,考察增强后语音的可懂度。实验结果表明,语音增强算法并不能改进语音的可懂度水平;通过分析具体的错误原因,发现听辩错误主要来自于音素错误,与声调关系不大;而且,同英文的辨识结果相比,一些增强算法对于中、英文可懂度影响差异显著。      

一种增强带噪语音可懂度的新算法

一般的语音增强算法在强噪声环境中只能提高信噪比,不能提高可懂度。本文提出用可调节白噪声代替信号中非语音部分的语音可懂度增强处理新算法。实验证明此方法能明显改善强噪声时的语音可懂度,能对低至-10dB的带噪语音信号进行有效的可懂度增强。     

一种面向声音变换的参数化模型

在源滤波器模型的基础上,利用统计学习方法,建立了一种面向声音变换的混合参数化模型。该模型包括浊音声学模型、清音声学模型和韵律补偿模型三部分。基于线性预测分析和mel倒谱分析的浊音声学模型,刻画了说话人声腔的共振特性。基于线性预测分析和噪声源分析的清音声学模型,反映了说话人发清音的特点。基于统计学习方法的韵律补偿模型描述了音高、能量与时长等分布特性。在该混合参数化模型的基础上,提出了一个声音变换算法,并将其应用到汉语音节的变换问题上。实验结果表明,对清浊音和韵律特性分别建模的变换算法能够提高重建语音的清晰度和可懂度,缩小重建语音与目标语音之间的感知距离,使重建语音具有目标说话人的韵律特征．     

汉语塞音选择性特征自动萃取的小波变换方法

本文提出了汉语语音导引特征的概念，讨论了语音导引特征在自动语音识别中用于导引匹配搜索的运用方式和重要作用；提出并设计了汉语塞音选择性特征自动萃取的小波变换方法和数字信号处理算法。本文方法和算法从声学信号处理和检测的角度，将汉语清辅音声波输入信号自动分为塞音子集ＢＤＧ：｛ｂ，ｄ，ｇ｝、塞音子集ＺＺＨＪＧＰＴｃＣＨＱＫ：｛ｚ，ｚｈ，ｊ，ｇ，ｐ，ｔ，ｃ，ｃｈ，ｑ，ｋ｝和擦音集ＦｓＳＨｈＸ：｛ｆ，ｓ，ｓｈ，ｘ，ｈ｝；对输入的合清辅音的音节，计算检测并输出汉语自动语音识别系统可以利用的清辅音类属标记ｂ．ｄ．ｇ、ＳＴＯＰ／ＢＤ和ｆ．ｓ．ｓｈ．ｘ．ｈ以及它们的音段起始时标；从声学信息计算检测的角度为汉语自动语音识别系统提供一种新的“从粗到细”的辅助匹配结构。算法可用性模拟实验采用实际语音的数据库数据，以手工标注信息作为自动检测分类正确与否的对比标准。对１２６７个汉语全音节中，总数９１３个待分类清辅音的初步分类结果表明：正确分类率分别为ｂ．ｄ．ｇ：９６．１％，ＳＴＯＰ／ＢＤ：９５．１％和ｆ．ｓ．ｓｈ．ｘ．ｈ：８９．０％，总体平均正确分类率为９３．６％。     

基于干扰协方差矩阵重构的恒定束宽鲁棒自适应波束形成

针对宽带波束形成中的恒定束宽波束响应优化设计问题与鲁棒性问题展开研究。首先,提出一种基于相位补偿的恒定束宽全局优化设计方法,通过对阵列流形向量进行相位补偿来设计恒定束宽波束,与现有的一些方法相比,该方法不仅能获得全局最优解,而且物理实现简单。同时,还提出一种基于协方差矩阵重构的鲁棒自适应宽带波束形成算法。该算法采用Capon估计器估计样本数据的空间一频率谱密度函数,然后对期望信号波达方向之外的角度区间进行积分来重构干扰加噪声协方差矩阵,最后利用重构的协方差矩阵设计自适应波束形成器权系数。该波束形成器设计问题被表述成凸优化问题求解。仿真结果表明,在整个输入信噪比范围内,该算法几乎都能获得接近理想值的输出信干噪比。     

Effects of the salience of pitch and periodicity information on the intelligibility of four-channel vocoded speech: implications for cochlear implants

Recent simulations of continuous interleaved sampling (CIS) cochlear implant speech processors have used acoustic stimulation that provides only weak cues to pitch, periodicity, and aperiodicity, although these are regarded as important perceptual factors of speech. Four-channel vocoders simulating CIS processors have been constructed, in which the salience of speech-derived periodicity and pitch information was manipulated. The highest salience of pitch and periodicity was provided by an explicit encoding, using a pulse carrier following fundamental frequency for voiced speech, and a noise carrier during voiceless speech. Other processors included noise-excited vocoders with envelope cutoff frequencies of 32 and 400 Hz. The use of a pulse carrier following fundamental frequency gave substantially higher performance in identification of frequency glides than did vocoders using envelope-modulated noise carriers. The perception of consonant voicing information was improved by processors that preserved periodicity, and connected discourse tracking rates were slightly faster with noise carriers modulated by envelopes with a cutoff frequency of 400 Hz compared to 32 Hz. However, consonant and vowel identification, sentence intelligibility, and connected discourse tracking rates were generally similar through all of the processors. For these speech tasks, pitch and periodicity beyond the weak information available from 400 Hz envelope-modulated noise did not contribute substantially to performance.     

Differential contribution of envelope fluctuations across frequency to consonant identification in quiet

Two experiments investigated the effects of critical bandwidth and frequency region on the use of temporal envelope cues for speech. In both experiments, spectral details were reduced using vocoder processing. In experiment 1, consonant identification scores were measured in a condition for which the cutoff frequency of the envelope extractor was half the critical bandwidth (HCB) of the auditory filters centered on each analysis band. Results showed that performance is similar to those obtained in conditions for which the envelope cutoff was set to 160 Hz or above. Experiment 2 evaluated the impact of setting the cutoff frequency of the envelope extractor to values of 4, 8, and 16 Hz or to HCB in one or two contiguous bands for an eight-band vocoder. The cutoff was set to 16 Hz for all the other bands. Overall, consonant identification was not affected by removing envelope fluctuations above 4 Hz in the low- and high-frequency bands. In contrast, speech intelligibility decreased as the cutoff frequency was decreased in the midfrequency region from 16 to 4 Hz. The behavioral results were fairly consistent with a physical analysis of the stimuli, suggesting that clearly measurable envelope fluctuations cannot be attenuated without affecting speech intelligibility.     

Compression and expansion of the temporal envelope: evaluation of speech intelligibility and sound quality.

Sensorineural hearing loss is accompanied by loudness recruitment, a steeper-than-normal rise of perceived loudness with presentation level. To compensate for this abnormality, amplitude compression is often applied (e.g., in a hearing aid). Alternatively, since speech intelligibility has been modeled as the perception of fast energy fluctuations, enlarging these (by means of expansion) may improve speech intelligibility. Still, even if these signal-processing techniques prove useful in terms of speech intelligibility, practical application might be hindered by unacceptably low sound quality. Therefore, both speech intelligibility and sound quality were evaluated for syllabic compression and expansion of the temporal envelope. Speech intelligibility was evaluated with an adaptive procedure, based on short everyday sentences either in noise or with a competing speaker. Sound quality was measured by means of a rating-scale procedure, for both speech and music. In a systematic setup, both the ratio of compression or expansion and the number of independent processing bands were varied. Individual hearing thresholds were compensated for by a listener-specific filter and amplification. Both listeners with normal hearing and listeners with sensorineural hearing impairment participated as paid volunteers. The results show that, on average, both compression and expansion fail to show better speech intelligibility or sound quality than linear amplification.     

The performance of different synthesis signals in acoustic models of cochlear implants

Synthesis (carrier) signals in acoustic models embody assumptions about perception of auditory electric stimulation. This study compared speech intelligibility of consonants and vowels processed through a set of nine acoustic models that used Spectral Peak (SPEAK) and Advanced Combination Encoder (ACE)-like speech processing, using synthesis signals which were representative of signals used previously in acoustic models as well as two new ones. Performance of the synthesis signals was determined in terms of correspondence with cochlear implant (CI) listener results for 12 attributes of phoneme perception (consonant and vowel recognition; F1, F2, and duration information transmission for vowels; voicing, manner, place of articulation, affrication, burst, nasality, and amplitude envelope information transmission for consonants) using four measures of performance. Modulated synthesis signals produced the best correspondence with CI consonant intelligibility, while sinusoids, narrow noise bands, and varying noise bands produced the best correspondence with CI vowel intelligibility. The signals that performed best overall (in terms of correspondence with both vowel and consonant attributes) were modulated and unmodulated noise bands of varying bandwidth that corresponded to a linearly varying excitation width of 0.4 mm at the apical to 8 mm at the basal channels.     

A comparative intelligibility study of single-microphone noise reduction algorithms

The evaluation of intelligibility of noise reduction algorithms is reported. IEEE sentences and consonants were corrupted by four types of noise including babble, car, street and train at two signal-to-noise ratio levels (0 and 5 dB), and then processed by eight speech enhancement methods encompassing four classes of algorithms: spectral subtractive, sub-space, statistical model based and Wiener-type algorithms. The enhanced speech was presented to normal-hearing listeners for identification. With the exception of a single noise condition, no algorithm produced significant improvements in speech intelligibility. Information transmission analysis of the consonant confusion matrices indicated that no algorithm improved significantly the place feature score, significantly, which is critically important for speech recognition. The algorithms which were found in previous studies to perform the best in terms of overall quality, were not the same algorithms that performed the best in terms of speech intelligibility. The subspace algorithm, for instance, was previously found to perform the worst in terms of overall quality, but performed well in the present study in terms of preserving speech intelligibility. Overall, the analysis of consonant confusion matrices suggests that in order for noise reduction algorithms to improve speech intelligibility, they need to improve the place and manner feature scores.     

The effects of the addition of low-level, low-noise noise on the intelligibility of sentences processed to remove temporal envelope information

The intelligibility of sentences processed to remove temporal envelope information, as far as possible, was assessed. Sentences were filtered into N analysis channels, and each channel signal was divided by its Hilbert envelope to remove envelope information but leave temporal fine structure (TFS) intact. Channel signals were combined to give TFS speech. The effect of adding low-level low-noise noise (LNN) to each channel signal before processing was assessed. The addition of LNN reduced the amplification of low-level signal portions that contained large excursions in instantaneous frequency, and improved the intelligibility of simple TFS speech sentences, but not more complex sentences. It also reduced the time needed to reach a stable level of performance. The recovery of envelope cues by peripheral auditory filtering was investigated by measuring the intelligibility of 'recovered-envelope speech', formed by filtering TFS speech with an array of simulated auditory filters, and using the envelopes at the output of these filters to modulate sinusoids with frequencies equal to the filter center frequencies (i.e., tone vocoding). The intelligibility of TFS speech and recovered-envelope speech fell as N increased, although TFS speech was still highly intelligible for values of N for which the intelligibility of recovered-envelope speech was low.     

Predicting speech intelligibility based on the signal-to-noise envelope power ratio after modulation-frequency selective processing

A model for predicting the intelligibility of processed noisy speech is proposed. The speech-based envelope power spectrum model has a similar structure as the model of Ewert and Dau [(2000). J. Acoust. Soc. Am. 108, 1181-1196], developed to account for modulation detection and masking data. The model estimates the speech-to-noise envelope power ratio, SNR(env), at the output of a modulation filterbank and relates this metric to speech intelligibility using the concept of an ideal observer. Predictions were compared to data on the intelligibility of speech presented in stationary speech-shaped noise. The model was further tested in conditions with noisy speech subjected to reverberation and spectral subtraction. Good agreement between predictions and data was found in all cases. For spectral subtraction, an analysis of the model's internal representation of the stimuli revealed that the predicted decrease of intelligibility was caused by the estimated noise envelope power exceeding that of the speech. The classical concept of the speech transmission index fails in this condition. The results strongly suggest that the signal-to-noise ratio at the output of a modulation frequency selective process provides a key measure of speech intelligibility.     

Relative importance of temporal information in various frequency regions for consonant identification in quiet and in noise

The relative importance of temporal information in broad spectral regions for consonant identification was assessed in normal-hearing listeners. For the purpose of forcing listeners to use primarily temporal-envelope cues, speech sounds were spectrally degraded using four-noise-band vocoder processing Frequency-weighting functions were determined using two methods. The first method consisted of measuring the intelligibility of speech with a hole in the spectrum either in quiet or in noise. The second method consisted of correlating performance with the randomly and independently varied signal-to-noise ratio within each band. Results demonstrated that all bands contributed equally to consonant identification when presented in quiet. In noise, however, both methods indicated that listeners consistently placed relatively more weight upon the highest frequency band. It is proposed that the explanation for the difference in results between quiet and noise relates to the shape of the modulation spectra in adjacent frequency bands. Overall, the results suggest that normal-hearing listeners use a common listening strategy in a given condition. However, this strategy may be influenced by the competing sounds, and thus may vary according to the context. Some implications of the results for cochlear implantees and hearing-impaired listeners are discussed.     

Intelligibility of average talkers in typical listening environments

Intelligibility of conversationally produced speech for normal hearing listeners was studied for three male and three female talkers. Four typical listening environments were used. These simulated a quiet living room, a classroom, and social events in two settings with different reverberation characteristics. For each talker, overall intelligibility and intelligibility for vowels, consonant voicing, consonant continuance, and consonant place were quantified using the speech pattern contrast (SPAC) test. Results indicated that significant intelligibility differences are observed among normal talkers even in listening environments that permit essentially full intelligibility for everyday conversations. On the whole, talkers maintained their relative intelligibility across the four environments, although there was one exception which suggested that some voices may be particularly susceptible to degradation due to reverberation. Consonant place was the most poorly perceived feature, followed by continuance, voicing, and vowel intelligibility. However, there were numerous significant interactions between talkers and speech features, indicating that a talker of average overall intelligibility may produce certain speech features with intelligibility that is considerably higher or lower than average. Neither long-term rms speech spectrum nor articulation rate was found to be an adequate single criterion for selecting a talker of average intelligibility. Ultimately, an average talker was chosen on the basis of four speech contrasts: initial consonant place, and final consonant place, voicing, and continuance.