• Journal of the Institute of Electronics and Information Engineers
• /
• v.51 no.5
• /
• pp.188-194
• /
• 2014

The feature extraction mechanism of the CASA(Computational Auditory Scene Analysis) system uses time continuity and frequency channel similarity to compose a correlogram of auditory elements. In segmentation, we compose a binary mask by using cross-correlation function, mask 1(speech) has the same periodicity and synchronization. However, when there is delay between autocorrelation signals with the same periodicity, it is determined as a speech, which is considered to be a drawback. In this paper, we proposed an algorithm to improve discrimination of channel similarity using Weighted Cross-correlation in segmentation. We conducted experiments to evaluate the speech segregation performance of the CASA system in background noise(siren, machine, white, car, crowd) environments by changing SNR 5dB and 0dB. In this paper, we compared the proposed algorithm to the conventional algorithm. The performance of the proposed algorithm has been improved as following: improvement of 2.75dB at SNR 5dB and 4.84dB at SNR 0dB for background noise environment.

• Speech Sciences
• /
• v.1
• /
• pp.153-200
• /
• 1997

In languages such as Japanese, it is very common to observe that short peripheral vowel are completely voiceless when surrounded by voiceless consonants. This phenomenon has been known as Montreal French, Shanghai Chinese, Greek, and Korean. Traditionally this phenomenon has been described as a phonological rule that either categorically deletes the vowel or changes the [+voice] feature of the vowel to [-voice]. This analysis was supported by Sawashima (1971) and Hirose (1971)'s observation that there are two distinct EMG patterns for voiced and devoiced vowel in Japanese. Close examination of the phonetic evidence based on acoustic data, however, shows that these phonological characterizations are not tenable (Jun & Beckman 1993, 1994). In this paper, we examined the vowel devoicing phenomenon in Korean using data from ENG fiberscopic and acoustic recorders of 100 sentences produced by one Korean speaker. The results show that there is variability in the 'degree of devoicing' in both acoustic and EMG signals, and in the patterns of glottal closing and opening across different devoiced tokens. There seems to be no categorical difference between devoiced and voiced tokens, for either EMG activity events or glottal patterns. All of these observations support the notion that vowel devoicing in Korean can not be described as the result of the application of a phonological rule. Rather, devoicing seems to be a highly variable 'phonetic' process, a more or less subtle variation in the specification of such phonetic metrics as degree and timing of glottal opening, or of associated subglottal pressure or intra-oral airflow associated with concurrent tone and stricture specifications. Some of token-pair comparisons are amenable to an explanation in terms of gestural overlap and undershoot. However, the effect of gestural timing on vocal fold state seems to be a highly nonlinear function of the interaction among specifications for the relative timing of glottal adduction and abduction gestures, of the amplitudes of the overlapped gestures, of aerodynamic conditions created by concurrent oral tonal gestures, and so on. In summary, to understand devoicing, it will be necessary to examine its effect on phonetic representation of events in many parts of the vocal tracts, and at many stages of the speech chain between the motor intent and the acoustic signal that reaches the hearer's ear.

• MALSORI
• /
• no.64
• /
• pp.77-88
• /
• 2007

Nowadays, mel-frequency cesptral coefficients (MFCCs) and Gaussian mixture models (GMMs) are used for the pathological voice detection. This paper suggests a method to improve the performance of the pathological/normal voice classification based on the MFCC-based GMM. We analyze the characteristics of the mel frequency-based filterbank energies using the fisher discriminant ratio (FDR). And the feature vectors through the linear discriminant analysis (LDA) transformation of the filterbank energies (FBE) and the MFCCs are implemented. An accuracy is measured by the GMM classifier. This paper shows that the FBE LDA-based GMM is a sufficiently distinct method for the pathological/normal voice classification, with a 96.6% classification performance rate. The proposed method shows better performance than the MFCC-based GMM with noticeable improvement of 54.05% in terms of error reduction.

• The Journal of the Acoustical Society of Korea
• /
• v.16 no.6
• /
• pp.25-35
• /
• 1997

Most of speech analysis methods developed up to date are based on second order statistics, and one of the biggest drawback of these methods is that they show dramatical performance degradation in noisy environments. On the contrary, the methods using higher order statistics(HOS), which has the property of suppressing Gaussian noise, enable robust feature extraction in noisy environments. In this paper we propose a text-independent speaker identification system using higher order statistics and compare its performance with that using the conventional second-order-statistics-based method in both white and colored noise environments. The proposed speaker identification system is based on the vector quantization approach, and employs HOS-based voiced/unvoiced detector in order to extract feature parameters for voiced speech only, which has non-Gaussian distribution and is known to contain most of speaker-specific characteristics. Experimental results using 50 speaker's database show that higher-order-statistics-based method gives a better identificaiton performance than the conventional second-order-statistics-based method in noisy environments.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.17 no.9
• /
• pp.993-1000
• /
• 1992

Recently, many researchers have found that the speaker recognition rate is high when they perform the speaker recognition using statistical processing method of orthogonal parameter, which are derived from the analysis of speech signal and contain much of the speaker's identity. This method, however, has problems caused by vocalization speed or time varying feature of speed. Thus, to solve these problems, this paper proposes two methods of speaker recognition which combine DTW algorithm with the method using orthogonal parameters extracted from $Karthumem-Lo\'{e}ve$ Transform method which applies orthogonal parameters as feature vector to ETW algorithm and the other is the method which applies orthogonal parameters to the optimal path. In addition, we compare speaker recognition rate obtained from the proposed two method with that from the conventional method of statistical process of orthogonal parameters. Orthogonal parameters used in this paper are derived from both linear prediction coefficients and partial correlation coefficients of speech signal.

• The Journal of the Acoustical Society of Korea
• /
• v.35 no.6
• /
• pp.501-509
• /
• 2016

We propose a method to improve the robustness of speaker verification on short test utterances. The accuracy of the state-of-the-art i-vector/probabilistic linear discriminant analysis systems can be degraded when testing utterance durations are short. The proposed method compensates for utterance variations of short test feature vectors using deep neural networks. We design three different types of DNN (Deep Neural Network) structures which are trained with different target output vectors. Each DNN is trained to minimize the discrepancy between the feed-forwarded output of a given short utterance feature and its original long utterance feature. We use short 2-10 s condition of the NIST (National Institute of Standards Technology, U.S.) 2008 SRE (Speaker Recognition Evaluation) corpus to evaluate the method. The experimental results show that the proposed method reduces the minimum detection cost relative to the baseline system.

• Speech Sciences
• /
• v.14 no.4
• /
• pp.99-114
• /
• 2007

This study attempted to acoustically analyze the diphthongs of native Jeju speakers in their 70s, 50s, and 20s to observe their diphthong inventories 96 subjects participated in the recording using a set of picture cards. Results show that Jeju dialect speakers in their 70s distinguish /we/ and $/w{\varepsilon}/$, /yo/ and /yc/, but have difficulty pronouncing /ye/, $/y{\varepsilon}/$, and /iy/ correctly. It is interesting to find that the diphthong inventory of Jeju dialect speakers in their 50s are in transitional stage. They share the diphthong inventories of those people in their 70's containing /we/ and $/w{\varepsilon}/$, and /yo/ and /yc/, but they do not produce them as clearly as the latter age group. The former age group also share the common feature with those people in their 20's because both age groups show the same pronunciation of /iy/. The youngest group seldom pronounce /we/ and $/w{\varepsilon}/$, /yo/ and /yc/ correctly unlike native Jeju speakers in their 70s, but they can easily pronounce the diphthong /iy/ like standard Korean speakers.

• The Journal of the Acoustical Society of Korea
• /
• v.32 no.2
• /
• pp.167-173
• /
• 2013

This work examines classification of diphthongs, as part of a distinctive feature-based speech recognition system. Acoustic measurements related to the vocal tract and the voice source are examined, and analysis of variance (ANOVA) results show that vowel duration, energy trajectory, and formant variation are significant. A balanced error rate of 17.8% is obtained for 2-way diphthong classification on the TIMIT database, and error rates of 32.9%, 29.9%, and 20.2% are obtained for /aw/, /ay/, and /oy/, for 4-way classification, respectively. Adding the acoustic features to widely used Mel-frequency cepstral coefficients also improves classification.

• Cross-Cultural Studies
• /
• v.28
• /
• pp.115-139
• /
• 2012

Interval or duration of segments, syllables, words and phrases is an important acoustic feature which influences the naturalness of speech. A number of cross-sectional studies regarding acoustic characteristics of children's speech development found that intervals of segments, syllables, words and phrases tend to change with the growing age. One hypothesis assumed that decreases in intervals would be greater when children were younger and smaller decreases in intervals when older (Thelen,1991), it has been supported by quite a number of researches on the basis of cross-sectional studies (Tingley & Allen,1975; Kent & Forner,1980; Chermak & Schneiderman, 1986), but the other hypothesis predicted that decreases in intervals would be smaller when children were younger and greater decreases in intervals when older (Smith, Kenney & Hussain, 1996). Researchers seem to come up with conflicting postulations and inconsistent results about the change trends concerning intervals of segments, syllables, words and phrases, leaving it as an issue unresolved. Most acoustic investigations of children's speech production have been conducted via cross-sectional designs, which involves studying several groups of children. So far, there are only a few longitudinal studies. This issue needs more longitudinal investigations; moreover, the acoustic measures of the intervals of child speech are hardly available. All former studies focus on word stages excluding the babbling stages especially the canonical babbling stage, but we need to find out when concrete changes of intervals begin to occur and what causes the changes. Therefore, we conducted an acoustic study of interval characteristics of segments and words concerning Canonical Babble ( CB) and early speech in an infant aged from 0;9 to 2;4 acquiring Mandarin Chinese. The current research addresses the following two questions: 1. Whether decreases in interval would be greater when children were younger and smaller when they were older or vice versa? 2. Whether the child speech concerning the acoustic features of interval drifts in the direction of the language they are exposed to? The female infant whose L1 was Southern Mandarin living in Changsha was audio- and video-taped at her home for about one hour almost on a weekly basis during her age range from 0;9 to 2;4 under natural observation by us investigators. The recordings were digitized. Parts of the digitized material were labeled. All the repetitions were excluded. The utterances were extracted from 44 sessions ranging from 30 minutes to one hour. The utterances were divided into segments as well as syllable-sized units. Age stages are 0;9-1;0,1;1-1;5, 1;6-2;0, 2;1-2;4. The subject was a monolingual normal child from parents with a good education. The infant was audio-and video-taped in her home almost every week. The data were digitized, segments and syllables from 44 sessions spanning the transition from babble to speech were transcribed in narrow IPA and coded for analysis. Babble was coded from age 0;9-1;0, and words were coded from 1;0 to 2;4, the data has been checked by two professionally trained persons who majored in phonetics. The present investigation is a longitudinal analysis of some temporal characteristics of the child speech during the age periods of 0;9-1;0, 1;1-1;5, 1;6-2;0, 2;1-2;4. The answer to Research Question 1 is that our results are in agreement with neither of the hypotheses. One hypothesis assumed that decreases in intervals would be greater when children were younger and smaller decreases in intervals when older (Thelen,1991); but the other hypothesis predicted that decreases in intervals would be smaller when children were younger and greater decreases in intervals when older (Smith, Kenney & Hussain, 1996). On the whole, there is a tendency of decrease in segmental and syllabic duration with the growing age, but the changes are not drastic and abrupt. For example, /a/ after /k/ in Table 1 has greater decrease during 1;1-1;5, while /a/ after /p/, /t/ and /w/ has greater decrease during 2;1-2;4. /ka/ has greater decrease during 1;1-1;5, while /ta/ and /na/ has greater decrease during 2;1-2;4.Across the age periods, interval change experiences lots of fluctuation all the time. The answer to Research Question 2 is yes. Babbling stage is a period in which the children's acoustic features of intervals of segments, syllables, words and phrases is shifted in the direction of the language to be learned, babbling and children's speech emergence is greatly influenced by ambient language. The phonetic changes in terms of duration would go on until as late as 10-12 years of age before reaching adult-like levels. Definitely, with the increase of exposure to ambient language, the variation would be less and less until they attain the adult-like competence. Via the analysis of the SPSS 15.0, the decrease of segmental and syllabic intervals across the four age periods proves to be of no significant difference (p>0.05). It means that the change of segmental and syllabic intervals is continuous. It reveals that the process of child speech development is gradual and cumulative.

• The Journal of the Acoustical Society of Korea
• /
• v.22 no.7
• /
• pp.519-527
• /
• 2003

Speech is much influenced by the existence of outliers which are introduced by such an unexpected happenings as additive background noise, change of speaker's utterance pattern and voice detection errors. These kinds of outliers may result in severe degradation of speaker recognition performance. In this paper, we proposed the GMM based on robust principal component analysis (RPCA-GMM) using M-estimation to solve the problems of both ouliers and high dimensionality of training feature vectors in speaker identification. Firstly, a new feature vector with reduced dimension is obtained by robust PCA obtained from M-estimation. The robust PCA transforms the original dimensional feature vector onto the reduced dimensional linear subspace that is spanned by the leading eigenvectors of the covariance matrix of feature vector. Secondly, the GMM with diagonal covariance matrix is obtained from these transformed feature vectors. We peformed speaker identification experiments to show the effectiveness of the proposed method. We compared the proposed method (RPCA-GMM) with transformed feature vectors to the PCA and the conventional GMM with diagonal matrix. Whenever the portion of outliers increases by every 2%, the proposed method maintains almost same speaker identification rate with 0.03% of little degradation, while the conventional GMM and the PCA shows much degradation of that by 0.65% and 0.55%, respectively This means that our method is more robust to the existence of outlier.