• The Journal of the Acoustical Society of Korea
• /
• v.38 no.6
• /
• pp.703-709
• /
• 2019

This paper proposes speech recognition systems employing Convolutional Neural Network (CNN) and Long Short Term Memory (LSTM) structures combined with Hidden Markov Moldel (HMM) to effectively recognize the speech of VeloPharyngeal Insufficiency (VPI) patients, and compares the recognition performance of the systems to the Gaussian Mixture Model (GMM-HMM) and fully-connected Deep Neural Network (DNNHMM) based speech recognition systems. In this paper, the initial model is trained using normal speakers' speech and simulated VPI speech is used for generating a prior model for speaker adaptation. For VPI speaker adaptation, selected layers are trained in the CNN-HMM based model, and dropout regulatory technique is applied in the LSTM-HMM based model, showing 3.68 % improvement in recognition accuracy. The experimental results demonstrate that the proposed LSTM-HMM-based speech recognition system is effective for VPI speech with small-sized speech data, compared to conventional GMM-HMM and fully-connected DNN-HMM system.

• The Transactions of the Korean Institute of Electrical Engineers
• /
• v.40 no.8
• /
• pp.782-791
• /
• 1991

In the fields of pattern recognition such as speech recognition, several new techniques using Artifical Neural network Models have been proposed and implemented. In particular, the Multilayer Perception Model has been shown to be effective in static speech pattern recognition. But speech has dynamic or temporal characteristics and the most important point in implementing speech recognition systems using Artificial Neural Network Models for continuous speech is the learning of dynamic characteristics and the distributed cues and contextual effects that result from temporal characteristics. But Recurrent Multilayer Perceptron Model is known to be able to learn sequence of pattern. In this paper, the results of applying the Recurrent Model which has possibilities of learning tedmporal characteristics of speech to phoneme recognition is presented. The test data consist of 144 Vowel+ Consonant + Vowel speech chains made up of 4 Korean monothongs and 9 Korean plosive consonants. The input parameters of Artificial Neural Network model used are the FFT coefficients, residual error and zero crossing rates. The Baseline model showed a recognition rate of 91% for volwels and 71% for plosive consonants of one male speaker. We obtained better recognition rates from various other experiments compared to the existing multilayer perceptron model, thus showed the recurrent model to be better suited to speech recognition. And the possibility of using Recurrent Models for speech recognition was experimented by changing the configuration of this baseline model.

• Journal of Korea Multimedia Society
• /
• v.7 no.11
• /
• pp.1597-1609
• /
• 2004

A lot of lipreading system has been proposed to compensate the rate of speech recognition dropped in a noisy environment. Previous lipreading systems work on some specific conditions such as artificial lighting and predefined background color. In this paper, we propose a real-time lipreading system which allows the motion of a speaker and relaxes the restriction on the condition for color and lighting. The proposed system extracts face and lip region from input video sequence captured with a common PC camera and essential visual information in real-time. It recognizes utterance words by using the visual information in real-time. It uses the hue histogram model to extract face and lip region. It uses mean shift algorithm to track the face of a moving speaker. It uses PCA(Principal Component Analysis) to extract the visual information for learning and testing. Also, it uses HMM(Hidden Markov Model) as a recognition algorithm. The experimental results show that our system could get the recognition rate of 90% in case of speaker dependent lipreading and increase the rate of speech recognition up to 40～85% according to the noise level when it is combined with audio speech recognition.

• Speech Sciences
• /
• v.14 no.4
• /
• pp.157-167
• /
• 2007

In this paper, we implemented a single-chip speech recognizer using the TMS320C2000 DSPs. For this implementation, we had developed very small-sized speaker-dependent recognition engine based on dynamic time warping, which is especially suited for embedded systems where the system resources are severely limited. We carried out some optimizations including speed optimization by programming time-critical functions in assembly language, and code size optimization and effective memory allocation. For the TMS320F2801 DSP which has 12Kbyte SRAM and 32Kbyte flash ROM, the recognizer developed can recognize 10 commands. For the TMS320F2808 DSP which has 36Kbyte SRAM and 128Kbyte flash ROM, it has additional capability of outputting the speech sound corresponding to the recognition result. The speech sounds for response, which are captured when the user trains commands, are encoded using ADPCM and saved on flash ROM. The single-chip recognizer needs few parts except for a DSP itself and an OP amp for amplifying microphone output and anti-aliasing. Therefore, this recognizer may play a similar role to dedicated speech recognition chips.

• 제어로봇시스템학회:학술대회논문집
• /
• 1987.10b
• /
• pp.473-478
• /
• 1987

The purpose of this study is to design a phoneme-class recognition system for Korean language using filter-bank analysis and zero crossing rate method. First, the speech signals are separated in 16 bandpass filters to obtain short-time spectrum of speech signals, and digitized by 16-ch A/D converter. And then, with the set of features which extracted from patterns of ratios of each channel energy level to overall energy level, the decision rules are made for recognize unknown speech signal. In this experiment, the recognition rate was about 93.1 percent for 7 vowels under multitalker environment and 74.4 percent for 10 initial sounds at single speaker.

• Journal of Information Processing Systems
• /
• v.14 no.1
• /
• pp.218-227
• /
• 2018

Vocal effort detection is important for both robust speech recognition and speaker recognition. In this paper, the spectral information entropy feature which contains more salient information regarding the vocal effort level is firstly proposed. Then, the model fusion method based on complementary model is presented to recognize vocal effort level. Experiments are conducted on isolated words test set, and the results show the spectral information entropy has the best performance among the three kinds of features. Meanwhile, the recognition accuracy of all vocal effort levels reaches 81.6%. Thus, potential of the proposed method is demonstrated.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.16 no.1
• /
• pp.49-54
• /
• 2006

The recent fingerprint recognition system has unstable factors, such as copy of fingerprint patterns and hacking of fingerprint feature point, which mali cause significant system error. Thus, in this research, we used the fingerprint as the main recognition device and then implemented the multi-biometric recognition system in serial using the speech recognition which has been widely used recently. As a multi-biometric recognition system, once the speech is successfully recognized, the fingerprint recognition process is run. In addition, speaker-dependent DTW(Dynamic Time Warping) algorithm is used among existing speech recognition algorithms (VQ, DTW, HMM, NN) for effective real-time process while KSOM (Kohonen Self-Organizing feature Map) algorithm, which is the artificial intelligence method, is applied for the fingerprint recognition system because of its calculation amount. The experiment of multi-biometric recognition system implemented in this research showed 2 to $7\%$ lower FRR (False Rejection Ratio) than single recognition systems using each fingerprints or voice, but zero FAR (False Acceptance Ratio), which is the most important factor in the recognition system. Moreover, there is almost no difference in the recognition time(average 1.5 seconds) comparing with other existing single biometric recognition systems; therefore, it is proved that the multi-biometric recognition system implemented is more efficient security system than single recognition systems based on various experiments.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.20 no.4
• /
• pp.528-533
• /
• 2010

This paper studied the training methods less affected by the emotional variation for the development of the robust speech recognition system. For this purpose, the effect of emotional variation on the speech signal and the speech recognition system were studied using speech database containing various emotions. The performance of the speech recognition system trained by using the speech signal containing no emotion is deteriorated if the test speech signal contains the emotions because of the emotional difference between the test and training data. In this study, it is observed that vocal tract length of the speaker is affected by the emotional variation and this effect is one of the reasons that makes the performance of the speech recognition system worse. In this paper, a training method that cover the speech variations is proposed to develop the emotionally robust speech recognition system. Experimental results from the isolated word recognition using HMM showed that propose method reduced the error rate of the conventional recognition system by 28.4% when emotional test data was used.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.19 no.6
• /
• pp.773-778
• /
• 2009

This paper studied the training methods less affected by the emotional variation for the development of the robust speech recognition system. For this purpose, the effect of emotional variations on the speech signal were studied using speech database containing various emotions. The performance of the speech recognition system trained by using the speech signal containing no emotion is deteriorated if the test speech signal contains the emotions because of the emotional difference between the test and training data. In this study, it is observed that vocal tract length of the speaker is affected by the emotional variation and this effect is one of the reasons that makes the performance of the speech recognition system worse. In this paper, vocal tract normalization method is used to develop the robust speech recognition system for emotional variations. Experimental results from the isolated word recognition using HMM showed that the vocal tract normalization method reduced the error rate of the conventional recognition system by 41.9% when emotional test data was used.

• 제어로봇시스템학회:학술대회논문집
• /
• 2004.08a
• /
• pp.1291-1295
• /
• 2004

This paper proposed a new approach for endpoint detection of isolated speech, which proves to significantly improve the endpoint detection performance. The proposed algorithm relies on the root mean square energy (rms energy), zero crossing rate and spectral characteristics of the speech signal where the Euclidean distance measure is adopted using cepstral coefficients to accurately detect the endpoint of isolated speech. The algorithm offers better performance than traditional energy-based algorithm. The vocabulary for the experiment includes English digit from one to nine. These experimental results were conducted by 360 utterances from a male speaker. Experimental results show that the accuracy of the algorithm is quite acceptable. Moreover, the computation overload of this algorithm is low since the cepstral coefficients parameters will be used in feature extraction later of speech recognition procedure.