A probabilistic framework for landmark detection based on phonetic features for automatic speech recognition

A probabilistic framework for a landmark-based approach to speech recognition is presented for obtaining multiple landmark sequences in continuous speech. The landmark detection module uses as input acoustic parameters (APs) that capture the acoustic correlates of some of the manner-based phonetic features. The landmarks include stop bursts, vowel onsets, syllabic peaks and dips, fricative onsets and offsets, and sonorant consonant onsets and offsets. Binary classifiers of the manner phonetic features-syllabic, sonorant and continuant-are used for probabilistic detection of these landmarks. The probabilistic framework exploits two properties of the acoustic cues of phonetic features-(1) sufficiency of acoustic cues of a phonetic feature for a probabilistic decision on that feature and (2) invariance of the acoustic cues of a phonetic feature with respect to other phonetic features. Probabilistic landmark sequences are constrained using manner class pronunciation models for isolated word recognition with known vocabulary. The performance of the system is compared with (1) the same probabilistic system but with mel-frequency cepstral coefficients (MFCCs), (2) a hidden Markov model (HMM) based system using APs and (3) a HMM based system using MFCCs.     

Detection of speech landmarks: use of temporal information

Studies by Shannon et al. [Science, 270, 303-304 (1995)], Van Tasell et al. [J. Acoust. Soc. Am. 82, 1152-1161 (1987)], and others show that human listeners can understand important aspects of the speech signal when spectral shape has been significantly degraded. These experiments suggest that temporal information is particularly important in human speech perception when the speech signal is heavily degraded. In this study, a system is developed that extracts linguistically relevant temporal information that can be used in the front end of an automatic speech recognition system. The parameters targeted include energy onset and offsets (computed using an adaptive algorithm) and measures of periodic and aperiodic content; together these are used to find abrupt acoustic events which signify landmarks. Overall detection rates for strongly robust events, robust events, and weak events in a portion of the TIMIT test database are 98.9%, 94.7%, and 52.1%, respectively. Error rates increase by less than 5% when the speech signals are spectrally impoverished. Use of the four temporal parameters as the front end of a hidden Markov model (HMM)-based system for the automatic recognition of the manner classes "sonorant," "fricative," "stop," and "silence" results in the same recognition accuracy achieved when the standard 39 cepstral-based parameters are used, 70.1%. The combination of the temporal parameters and cepstral parameters results in an accuracy of 74.8%.     

Analysis of stop consonant production errors in developmentally dysphasic children

The speech production skills of 12 dysphasic children and of 12 normal children were compared. The dysphasic children were found to have significantly greater difficulty than the normal children in producing stop consonants. In addition, it was found that seven of the dysphasic children, who had difficulty in perceiving initial stop consonants, had greater difficulty in producing stop consonants than the remaining five dysphasic children who showed no such perceptual difficulty. A detailed phonetic analysis indicated that the dysphasic children seldom omitted stops or substituted nonstop for stop consonants. Instead, their errors were predominantly of voicing or place of articulation. Acoustic analyses suggested that the voicing errors were related to lack of precise control over the timing of speech events, specifically, voice onset time for initial stops and vowel duration preceding final stops. The number of voicing errors on final stops, however, was greater than expected on the basis of lack of differentiation of vowel duration alone. They appeared also to be related to a tendency in the dysphasic children to produce final stops with exaggerated aspiration. The possible relationship of poor timing control in speech production in these children and auditory temporal processing deficits in speech perception is discussed.     

Characteristics of listener sensitivity to talker-specific phonetic detail

Previous research shows that listeners are sensitive to talker differences in phonetic properties of speech, including voice-onset-time (VOT) in word-initial voiceless stop consonants, and that learning how a talker produces one voiceless stop transfers to another word with the same voiceless stop [Allen, J. S., and Miller, J. L. (2004). J. Acoust. Soc. Am. 115, 3171-3183]. The present experiments examined whether transfer extends to words that begin with different voiceless stops. During training, listeners heard two talkers produce a given voiceless-initial word (e.g., pain). VOTs were manipulated such that one talker produced the voiceless stop with relatively short VOTs and the other with relatively long VOTs. At test, listeners heard a short- and long-VOT variant of the same word (e.g., pain) or a word beginning with a different voiceless stop (e.g., cane or coal), and were asked to select which of the two VOT variants was most representative of a given talker. In all conditions, which variant was selected at test was in line with listeners' exposure during training, and the effect was equally strong for the novel word and the training word. These findings suggest that accommodating talker-specific phonetic detail does not require exposure to each individual phonetic segment.     

Analysis of acoustic parameters for consonant voicing classification in clean and telephone speech

This paper describes acoustic cues for classification of consonant voicing in a distinctive feature-based speech recognition system. Initial acoustic cues are selected by studying consonant production mechanisms. Spectral representations, band-limited energies, and correlation values, along with Mel-frequency cepstral coefficients features (MFCCs) are also examined. Analysis of variance is performed to assess relative significance of features. Overall, 82.2%, 80.6%, and 78.4% classification rates are obtained on the TIMIT database for stops, fricatives, and affricates, respectively. Combining acoustic parameters with MFCCs shows performance improvement in all cases. Also, performance in the NTIMIT telephone channel speech shows that acoustic parameters are more robust than MFCCs.     

Auditory/phonetic categorization with the Symbion multichannel cochlear implant

The phonetic identification ability of an individual (SS) who exhibits the best, or equal to the best, speech understanding of patients using the Symbion four-channel cochlear implant is described. It has been found that SS: (1) can use aspects of signal duration to form categories that are isomorphic with the phonetic categories established by listeners with normal auditory function; (2) can combine temporal and spectral cues in a normal fashion to form categories; (3) can use aspects of fricative noises to form categories that correspond to normal phonetic categories; (4) uses information from both F1 and higher formants in vowel identification; and (5) appears to identify stop consonant place of articulation on the basis of information provided by the center frequency of the burst and by the abruptness of frequency change following signal onset. SS has difficulty identifying stop consonants from the information provided by formant transitions and cannot differentially identify signals that have identical F1's and relatively low-frequency F2's. SS's performance suggests that simple speech processing strategies (filtering of the signal into four bands) and monopolar electrode design are viable options in the design of cochlear prostheses.     

联合深度编解码网络和时频掩蔽估计的单通道语音增强

提出了一种联合深度编解码神经网络和时频掩蔽估计的语音增强方法。该方法利用深度编解码网络估计时频掩蔽表示,并联合带噪语音的幅度谱学习带噪语音与纯净语音幅度谱之间的非线性映射关系。深度编解码网络采用卷积-反卷积网络结构。在编码端,利用卷积网络的局部感知特性,对带噪语音的时频域结构特征进行建模,提取语音特征,同时抑制背景噪声。在解码端,利用编码端提取到的语音特征逐层恢复局部细节信息并重构语音信号。同时,在编解码端对应层之间引入跳跃连接,以减少由于池化和全连接操作导致的低层细节信息丢失的问题。在TIMIT语音库和不完全匹配噪声集下进行仿真实验,实验结果表明,该方法可以有效抑制噪声,且能较好地恢复出语音细节成分。      

汉语综合资料库的设计

语言是人类最重要的交际工具，随着现代信息技术的发展，语言也是人与机器之间交际的有效工具．近年来世界各国纷纷建立本国的言语资料库作为言语科学研究和言语技术开发的基础．汉语综合资料库的语音材料有：汉语全部有调音节、数字串、单词、韵律特征材料，以及语言清晰度试验用音节表、词表、句表和有代表性的短文等．汉语综合资料库在语言学和语音学特征以及声学特征方面充分体现汉语的基本特点．首先要解决语料选取问题，考虑各种语言单位的使用频率，不仅要包括全部高频词，也要反映较全面的语音现象．数据库在结构上是开放的模块式的，同时配有灵活的数据库管理系统．     

Effects and modeling of phonetic and acoustic confusions in accented speech

Accented speech recognition is more challenging than standard speech recognition due to the effects of phonetic and acoustic confusions. Phonetic confusion in accented speech occurs when an expected phone is pronounced as a different one, which leads to erroneous recognition. Acoustic confusion occurs when the pronounced phone is found to lie acoustically between two baseform models and can be equally recognized as either one. We propose that it is necessary to analyze and model these confusions separately in order to improve accented speech recognition without degrading standard speech recognition. Since low phonetic confusion units in accented speech do not give rise to automatic speech recognition errors, we focus on analyzing and reducing phonetic and acoustic confusability under high phonetic confusion conditions. We propose using likelihood ratio test to measure phonetic confusion, and asymmetric acoustic distance to measure acoustic confusion. Only accent-specific phonetic units with low acoustic confusion are used in an augmented pronunciation dictionary, while phonetic units with high acoustic confusion are reconstructed using decision tree merging. Experimental results show that our approach is effective and superior to methods modeling phonetic confusion or acoustic confusion alone in accented speech, with a significant 5.7% absolute WER reduction, without degrading standard speech recognition.     

Contextual effects on vowel duration, closure duration, and the consonant/vowel ratio in speech production

Acoustic measurements were conducted to determine the degree to which vowel duration, closure duration, and their ratio distinguish voicing of word-final stop consonants across variations in sentential and phonetic environments. Subjects read CVC test words containing three different vowels and ending in stops of three different places of articulation. The test words were produced either in nonphrase-final or phrase-final position and in several local phonetic environments within each of these sentence positions. Our measurements revealed that vowel duration most consistently distinguished voicing categories for the test words. Closure duration failed to consistently distinguish voicing categories across the contextual variables manipulated, as did the ratio of closure and vowel duration. Our results suggest that vowel duration is the most reliable correlate of voicing for word-final stops in connected speech.     

Statistical modeling of speech Poincaré sections in combination of frequency analysis to improve speech recognition performance

This paper introduces a combinational feature extraction approach to improve speech recognition systems. The main idea is to simultaneously benefit from some features obtained from Poincare? section applied to speech reconstructed phase space (RPS) and typical Mel frequency cepstral coefficients (MFCCs) which have a proved role in speech recognition field. With an appropriate dimension, the reconstructed phase space of speech signal is assured to be topologically equivalent to the dynamics of the speech production system, and could therefore include information that may be absent in linear analysis approaches. Moreover, complicated systems such as speech production system can present cyclic and oscillatory patterns and Poincare? sections could be used as an effective tool in analysis of such trajectories. In this research, a statistical modeling approach based on Gaussian mixture models (GMMs) is applied to Poincare? sections of speech RPS. A final pruned feature set is obtained by applying an efficient feature selection approach to the combination of the parameters of the GMM model and MFCC-based features. A hidden Markov model-based speech recognition system and TIMIT speech database are used to evaluate the performance of the proposed feature set by conducting isolated and continuous speech recognition experiments. By the proposed feature set, 5.7% absolute isolated phoneme recognition improvement is obtained against only MFCC-based features.     

Individual talker differences in voice-onset-time

Individual talkers differ in the acoustic properties of their speech, and at least some of these differences are in acoustic properties relevant for phonetic perception. Recent findings from studies of speech perception have shown that listeners can exploit such differences to facilitate both the recognition of talkers' voices and the recognition of words spoken by familiar talkers. These findings motivate the current study, whose aim is to examine individual talker variation in a particular phonetically-relevant acoustic property, voice-onset-time (VOT). VOT is a temporal property that robustly specifies voicing in stop consonants. From the broad literature involving VOT, it appears that individual talkers differ from one another in their VOT productions. The current study confirmed this finding for eight talkers producing monosyllabic words beginning with voiceless stop consonants. Moreover, when differences in VOT due to variability in speaking rate across the talkers were factored out using hierarchical linear modeling, individual talkers still differed from one another in VOT, though these differences were attenuated. These findings provide evidence that VOT varies systematically from talker to talker and may therefore be one phonetically-relevant acoustic property underlying listeners' capacity to benefit from talker-specific experience.     

Acoustic properties of naturally produced clear speech at normal speaking rates

Sentences spoken "clearly" are significantly more intelligible than those spoken "conversationally" for hearing-impaired listeners in a variety of backgrounds [Picheny et al., J. Speech Hear. Res. 28, 96-103 (1985); Uchanski et al., ibid. 39, 494-509 (1996); Payton et al., J. Acoust. Soc. Am. 95, 1581-1592 (1994)]. While producing clear speech, however, talkers often reduce their speaking rate significantly [Picheny et al., J. Speech Hear. Res. 29, 434-446 (1986); Uchanski et al., ibid. 39, 494-509 (1996)]. Yet speaking slowly is not solely responsible for the intelligibility benefit of clear speech (over conversational speech), since a recent study [Krause and Braida, J. Acoust. Soc. Am. 112, 2165-2172 (2002)] showed that talkers can produce clear speech at normal rates with training. This finding suggests that clear speech has inherent acoustic properties, independent of rate, that contribute to improved intelligibility. Identifying these acoustic properties could lead to improved signal processing schemes for hearing aids. To gain insight into these acoustical properties, conversational and clear speech produced at normal speaking rates were analyzed at three levels of detail (global, phonological, and phonetic). Although results suggest that talkers may have employed different strategies to achieve clear speech at normal rates, two global-level properties were identified that appear likely to be linked to the improvements in intelligibility provided by clear/normal speech: increased energy in the 1000-3000-Hz range of long-term spectra and increased modulation depth of low frequency modulations of the intensity envelope. Other phonological and phonetic differences associated with clear/normal speech include changes in (1) frequency of stop burst releases, (2) VOT of word-initial voiceless stop consonants, and (3) short-term vowel spectra.     

Syllable structure and integration of voicing and manner of articulation information in labial consonant identification

Speech perception requires the integration of information from multiple phonetic and phonological dimensions. A sizable literature exists on the relationships between multiple phonetic dimensions and single phonological dimensions (e.g., spectral and temporal cues to stop consonant voicing). A much smaller body of work addresses relationships between phonological dimensions, and much of this has focused on sequences of phones. However, strong assumptions about the relevant set of acoustic cues and/or the (in)dependence between dimensions limit previous findings in important ways. Recent methodological developments in the general recognition theory framework enable tests of a number of these assumptions and provide a more complete model of distinct perceptual and decisional processes in speech sound identification. A hierarchical Bayesian Gaussian general recognition theory model was fit to data from two experiments investigating identification of English labial stop and fricative consonants in onset (syllable initial) and coda (syllable final) position. The results underscore the importance of distinguishing between conceptually distinct processing levels and indicate that, for individual subjects and at the group level, integration of phonological information is partially independent with respect to perception and that patterns of independence and interaction vary with syllable position.     

Effects of deafness on acoustic characteristics of American English tense/lax vowels in maternal speech to infants

Recent studies have demonstrated that mothers exaggerate phonetic properties of infant-directed (ID) speech. However, these studies focused on a single acoustic dimension (frequency), whereas speech sounds are composed of multiple acoustic cues. Moreover, little is known about how mothers adjust phonetic properties of speech to children with hearing loss. This study examined mothers' production of frequency and duration cues to the American English tense/lax vowel contrast in speech to profoundly deaf (N?=?14) and normal-hearing (N?=?14) infants, and to an adult experimenter. First and second formant frequencies and vowel duration of tense (/i/,?/u/) and lax (/I/,?/?/) vowels were measured. Results demonstrated that for both infant groups mothers hyperarticulated the acoustic vowel space and increased vowel duration in ID speech relative to adult-directed speech. Mean F2 values were decreased for the /u/ vowel and increased for the /I/ vowel, and vowel duration was longer for the /i/, /u/, and /I/ vowels in ID speech. However, neither acoustic cue differed in speech to hearing-impaired or normal-hearing infants. These results suggest that both formant frequencies and vowel duration that differentiate American English tense/lx vowel contrasts are modified in ID speech regardless of the hearing status of the addressee.     

Influence of following context on perception of the voiced-voiceless distinction in syllable-final stop consonants

This paper reports acoustic measurements and results from a series of perceptual experiments on the voiced-voiceless distinction for syllable-final stop consonants in absolute final position and in the context of a following syllable beginning with a different stop consonant. The focus is on temporal cues to the distinction, with vowel duration and silent closure duration as the primary and secondary dimensions, respectively. The main results are that adding a second syllable to a monosyllable increases the number of voiced stop consonant responses, as does shortening of the closure duration in disyllables. Both of these effects are consistent with temporal regularities in speech production: Vowel durations are shorter in the first syllable of disyllables than in monosyllables, and closure durations are shorter for voiced than for voiceless stops in disyllabic utterances of this type. While the perceptual effects thus may derive from two separate sources of tacit phonetic knowledge available to listeners, the data are also consistent with an interpretation in terms of a single effect; one of temporal proximity of following context.     

谐波显著度的基频提取方法

我们提出的谐波显著度的基频提取方法,目的是从语音信号中自动获取人声基频,该方法利用抑制因子计算出基频的谐波显著度谱,对各次谐波显著度加权求和之后进行基频轨迹跟踪确定语音的基频序列。在TIMIT掺噪数据集和音乐信息检索评测2005主旋律数据集上,谐波显著度方法的准确率分别达到了88.5%和73.3%,使倍频、半频错误相对降低了80%。实验表明,基于谐波显著度的基频提取方法增强了系统的抗噪性能以及抗倍半频错误的能力。      

On the perception of similarity among talkers

A listener who recognizes a talker notices characteristic attributes of the talker's speech despite the novelty of each utterance. Accounts of talker perception have often presumed that consistent aspects of an individual's speech, termed indexical properties, are ascribable to a talker's unique anatomy or consistent vocal posture distinct from acoustic correlates of phonetic contrasts. Accordingly, the perception of a talker is acknowledged to occur independently of the perception of a linguistic message. Alternatively, some studies suggest that attention to attributes of a talker includes indexical linguistic attributes conveyed in the articulation of consonants and vowels. This investigation sought direct evidence of attention to phonetic attributes of speech in perceiving talkers. Natural samples and sinewave replicas derived from them were used in three experiments assessing the perceptual properties of natural and sine-wave sentences; of temporally veridical and reversed natural and sine-wave sentences; and of an acoustic correlate of vocal tract scale to judgments of sine-wave talker similarity. The results revealed that the subjective similarity of individual talkers is preserved in the absence of natural vocal quality; and that local phonetic segmental attributes as well as global characteristics of speech can be exploited when listeners notice characteristics of talkers.