• The Journal of the Acoustical Society of Korea
• /
• v.22 no.3E
• /
• pp.124-132
• /
• 2003

We propose a new method for speech/non-speech classifiers based on concepts of the adaptive boosting (AdaBoost) algorithm in order to detect speech for robust speech recognition. The method uses a combination of simple base classifiers through the AdaBoost algorithm and a set of optimized speech features combined with spectral subtraction. The key benefits of this method are the simple implementation, low computational complexity and the avoidance of the over-fitting problem. We checked the validity of the method by comparing its performance with the speech/non-speech classifier used in a standard voice activity detector. For speech recognition purpose, additional performance improvements were achieved by the adoption of new features including speech band energies and MFCC-based spectral distortion. For the same false alarm rate, the method reduced 20-50% of miss errors.

• Phonetics and Speech Sciences
• /
• v.1 no.3
• /
• pp.155-163
• /
• 2009

This paper introduces an automatic voice classification system for the diagnosis of individual constitution based on Sasang Constitutional Medicine (SCM) in Traditional Korean Medicine (TKM). For the developing of this algorithm, we used the voices of 473 speakers and extracted a total of 144 speech features from the speech data consisting of five sustained vowels and one sentence. The classification system, based on a rule-based algorithm that is derived from a non parametric statistical method, presents binary negative decisions. In conclusion, 55.7% of the speech data were diagnosed by this system, of which 72.8% were correct negative decisions.

• Phonetics and Speech Sciences
• /
• v.6 no.1
• /
• pp.71-76
• /
• 2014

In this paper voice quality of normal speech is qualitatively classified by five components of breathy, creaky, rough, nasal, and thin/thick voice. To determine whether a correlation exists between a subjective measure of voice and an objective measure of voice, each voice is perceptually evaluated using the 1/2/3 scale by speech processing specialists and acoustically analyzed using speech analysis tools such as the Praat, MDVP, and VoiceSauce. The speech parameters include features related to speech source and vocal tract filter. Statistical analysis uses a two-independent-samples non-parametric test. Experimental results show that statistical analysis identified a significant correlation between the speech feature parameters and the components of voice quality.

• Phonetics and Speech Sciences
• /
• v.2 no.1
• /
• pp.77-85
• /
• 2010

This paper proposes an efficient feature vector processing technique to guard the Speech Emotion Recognition (SER) system against a variety of noises. In the proposed approach, emotional feature vectors are extracted from speech processed by comb filtering. Then, these extracts are used in a robust model construction based on feature vector classification. We modify conventional comb filtering by using speech presence probability to minimize drawbacks due to incorrect pitch estimation under background noise conditions. The modified comb filtering can correctly enhance the harmonics, which is an important factor used in SER. Feature vector classification technique categorizes feature vectors into either discriminative vectors or non-discriminative vectors based on a log-likelihood criterion. This method can successfully select the discriminative vectors while preserving correct emotional characteristics. Thus, robust emotion models can be constructed by only using such discriminative vectors. On SER experiment using an emotional speech corpus contaminated by various noises, our approach exhibited superior performance to the baseline system.

• Speech Sciences
• /
• v.8 no.2
• /
• pp.61-71
• /
• 2001

Speech material, which is collected from ARS(Automatic Response System), was analyzed and classified into disease and non-disease state. The material include 11 different kinds of diseases. Along with ARS speech, DAT(Digital Audio Tape) speech is collected in parallel to give the bench mark. To analyze speech material, analysis tools, which is developed local laboratory, are used to provide an improved and robust performance to the obtained parameters. To classify speech into disease and non-disease class, multi-layered neural network was used. Three different combinations of 3, 6, 12 parameters are tested to obtain the proper network size and to find the best performance. From the experiment, the classification rate of 92.5% was obtained.

• Phonetics and Speech Sciences
• /
• v.7 no.1
• /
• pp.117-124
• /
• 2015

This paper investigates the acoustic characteristics of English fricatives in the TIMIT corpus, with a special focus on the role of gender in rendering fricatives in American English. The TIMIT database includes 630 talkers and 2342 different sentences, comprising over five hours of speech. Acoustic analyses are conducted in the domain of spectral and temporal properties by treating gender as an independent factor. The results of acoustic analyses revealed that the most acoustic properties of voiceless sibilants turned out to be different between male and female speakers, but those of voiceless non-sibilants did not show differences. A classification experiment using linear discriminant analysis (LDA) revealed that 85.73% of voiceless fricatives are correctly classified. The sibilants are 88.61% correctly classified, whereas the non-sibilants are only 57.91% correctly classified. The majority of the errors are from the misclassification of /ɵ/ as [f]. The average accuracy of gender classification is 77.67%. Most of the inaccuracy results are from the classification of female speakers in non-sibilants. The results are accounted for by resorting to biological differences as well as macro-social factors. The paper contributes to the understanding of the role of gender in a large-scale speech corpus.

• Journal of the Korea Convergence Society
• /
• v.8 no.5
• /
• pp.13-17
• /
• 2017

Existing Speech feature extracting method in speech Signal, there are incorrect recognition rates due to incorrect speech which is not clear threshold value. In this article, the modeling method for improving speech recognition performance that combines the feature extraction for speech and silence characteristics normalized to the non-speech. The proposed method is minimized the noise affect, and speech recognition model are convergence of speech signal feature extraction to each speech frame and the silence feature normalization. Also, this method create the original speech signal with energy spectrum similar to entropy, therefore speech noise effects are to receive less of the noise. the performance values are improved in signal to noise ration by the silence feature normalization. We fixed speech and non speech classification standard value in cepstrum For th Performance analysis of the method presented in this paper is showed by comparing the results with CHMM HMM, the recognition rate was improved 2.7%p in the speech dependent and advanced 0.7%p in the speech independent.

• ETRI Journal
• /
• v.36 no.5
• /
• pp.714-720
• /
• 2014

In this paper, we propose the use of data-driven probabilistic utterance-level decision logic to improve Weighted Finite State Transducer (WFST)-based endpoint detection. In general, endpoint detection is dealt with using two cascaded decision processes. The first process is frame-level speech/non-speech classification based on statistical hypothesis testing, and the second process is a heuristic-knowledge-based utterance-level speech boundary decision. To handle these two processes within a unified framework, we propose a WFST-based approach. However, a WFST-based approach has the same limitations as conventional approaches in that the utterance-level decision is based on heuristic knowledge and the decision parameters are tuned sequentially. Therefore, to obtain decision knowledge from a speech corpus and optimize the parameters at the same time, we propose the use of data-driven probabilistic utterance-level decision logic. The proposed method reduces the average detection failure rate by about 14% for various noisy-speech corpora collected for an endpoint detection evaluation.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.20 no.9
• /
• pp.1673-1678
• /
• 2016

In the area of the speech recognition processing, background noises are caused the incorrect response to the speech input, therefore the speech recognition rates are decreased by the background noises. Accordingly, a more high level noise processing techniques are required since these kinds of noise countermeasures are not simple. Therefore, this paper proposes an algorithm to distinguish between the stationary background noises or non-stationary background noises and the speech signal having short time duration in the noisy environments. The proposed algorithm uses the characteristic parameter of the improved speech signal as an important measure in order to distinguish different types of the background noises and the speech signals. Next, this algorithm estimates various kinds of the background noises using a multi-layer perceptron neural network. In this experiment, it was experimentally clear the estimation of the background noises and the speech signals.

• Journal of the Korea Convergence Society
• /
• v.5 no.1
• /
• pp.35-41
• /
• 2014

An automatic speech recognition system is one of the popular research problems. There are many research groups working in this field for different language including Japanese. Japanese vowel recognition is one of important parts in the Japanese speech recognition system. The vowel classification system with the Mamdani fuzzy inference system was developed in this research. We tested our system on the blind test data set collected from one male native Japanese speaker and four male non-native Japanese speakers. All subjects in the blind test data set were not the same subjects in the training data set. We found out that the classification rate from the training data set is 95.0 %. In the speaker-independent experiments, the classification rate from the native speaker is around 70.0 %, whereas that from the non-native speakers is around 80.5 %.