다층회귀예측신경망의 음성인식성능에 관한 연구

안점영 한국정보통신학회 1999 한국정보통신학회논문지 Vol.3 No.2

4층구조의 다층퍼셉트론을 변형하여 3 종류의 다층회귀예측신경망을 구성하고, 예측차수, 두 은닉층의 뉴런개수, 연결세기의 초기치 및 전달함수 변화에 따른 각 망의 음성인식성능을 실험을 통해 각각 비교 분석한다. 실험결과에 의하면, 다층회귀신경망이 다층퍼셉트론에 비해 음성인식성능이 우수하다. 그리고 구조적으로는 상위은닉층의 출력을 하위은닉층으로 회귀할 때 인식성능이 가장 우수하며, 각 망 공히 상, 하위은닉층의 뉴런 10 혹은 15개, 예측차수 3 혹은 4차일 때 인식률이 양호하다. 학습시 연결세기의 초기치를 -0.5에서 0.5사이로 설정하고, 하위은닉층에서 단극성 시그모이드 전달함수를 사용할 때 인식성능이 더욱 향상된다. We devise the 3 models of Multilayered Recurrent Prediction Neural Network(MLRPNN), which are obtained by modifying the Multilayered Perceptron(MLP) with 4 layers. We experimentally study the speech recognition performance of 3 models by a comparative method, according to the variation of the prediction order, the number of neurons in two hidden layers, initial values of connecting weights and transfer function, respectively. By the experiment, the recognition performance of each MLRPNN is better than that of MLP. At the model that returns the output of the upper hidden layer to the lower hidden layer, the recognition performance shows the best value. All MLRPNNs, which have 10 or 15 neurons in the upper and lower hidden layer and is predicted by 3rd or 4th order, show the improved speech recognition rate. On learning, these MLRPNNs have a better recognition rate when we set the initial weights between -0.5 and 0.5, and use the unipolar sigmoid transfer function in the lower hidden layer.

음성 합성에서 성문파가 음질에 미치는 영향 ( Effect of Glottal Wave on teh Quality )

안점영,김명기 대한전자공학회 1985 대한전자공학회 학술대회 Vol.1985 No.1

한국어 CV단음절의 음소합성

안점영,김명기 한국통신학회 1986 韓國通信學會論文誌 Vol.11 No.2

子音 音素/ㄱ, ㄷ, ㅂ, ㅈ/과 이에 대응한 硬音, 激音 그리고 母音 音素/ㅏ, ㅓ, ㅗ, ㅜ, ㅣ/로 구성된 韓國語 CV單音節을 偏自己相關方式으로 分析하고, 分析된 parameter를 적절히 제어하여 音素合成方法으로 이들 音節을 合成하였다. 분석결과 자음길이는 激音일 때 제일 길고, 硬音이 가장 짧았으며 이 音들의 gain도 비슷한 변화를 나타내었다. 그리고 平音뒤의 모음 pitch 주기가 가장 길고, 硬音, 激音으로 바뀌면 pitch주기가 짧아졌다. 子音 音素는 激音의 길이와 gain을 제어하여 합성하고 母音 音素는 平音뒤에 오는 母音의 pitch와 길이를 제어하여 합성하였다. 子音과 母音 音素를 結合시켜 CV單音節을 合成하였다. 實驗結果 合成音質은 대체로 양호하였고, 韓國語 音聲의 音素合成에 필요한 規則作成의 可能性을 확인하였다. We analyzed Korean CV mono-syllables consisted of concatenation of consonants/k, t, p, g/, their fortis and rough sound and vowels/a, e, o, u, I/by the PARCOR technique, and then we synthesized those speech by means of the phoneme synthesis controlling the analyzed data. In the speech analysis, the duration of consonants decreases in the rough sound, the lenis and the fortis in turns. And also the gain of them decreases in the same tendency. The pitch period increases more and more in vowels following the rough sound, the fortis and the lenis in turns. We synthesized the lenis and the fortis by controlling the duration and the gain of the rough sound, and vowels following the fortis and the rough sound by controlling the pitch period and the duration of vowels following the lenis. As the results, the synthesized speech quality is good and we make certain it is possible to make a rule to the phonome synthesis in Korea speech.

音聲合成에서 聲門 MODEL이 音質에 미치는 影響

金明起,安點榮 東亞大學校 1985 東亞論叢 Vol.22 No.1

It is investigated how the pitch period, gain, and glottal wave influence the synthesized quality. Speech sounds pronounced by a male are analyzed by an autocorrelation technique, and pitch periods and gains are obtained. In the synthesis of speech sounds, the effects of variations of the pitch periods and gains on the speech quality are recognized in terms of waveforms and listening. There is little degradation in synthesized speech quality with some variation of pitch period, but much difference in the clearness and naturalness of speech with variations of gain pattern-approximations, However, it is possible to treat the variation of pitch period and that of gain respectively as piecewise linear approximations without degrading the speech quality. Five Korean vowels are resynthesized by using a simulated glottal wave with nonimpulse shape, and it is tried to find the effect of opening time and closing time of the glottal wave on the speech quality. Main results obtained are as follows: 1. It is confirmed that the glottal waves are different in waveshape depending on the kind of vowels /a, e, i, o, u/ by inverse filtering technique. 2. It is found that the synthesized speech quality is varied with the opening time and closing time of glottal wave.

2단 회귀신경망의 숫자음 인식에관한 연구

안점영 한국통신학회 2000 韓國通信學會論文誌 Vol.25 No.3

We compose the two-stage recurrent neural network that returns both signals of a hidden and an output layer to the hidden layer. It is tested on the basis of syllables for Korean spoken digit from /gong/to /gu. For these experiments, we adjust the neuron number of the hidden layer, the predictive order of input data and self-recurrent coefficient of the decision state layer. By the experimental results, the recognition rate of this neural network is between 91% and 97.5% in the speaker-dependent case and between 80.75% and 92% in the speaker-independent case. In the speaker-dependent case, this network shows an equivalent recognition performance to Jordan and Elman network but in the speaker-independent case, it does improved performance.

소어휘 단어단위의 음성인식 칩 설계

안점영,최영식 한국정보통신학회 2002 한국정보통신학회논문지 Vol.6 No.2

소어휘 단어단위의 음성을 인식할 수 있는 음성인식 칩을 설계하였다. 설계된 칩은 음성 신호의 시작과 끝점 검출 부분, LPC 켑스트럼 계수 추출 부분, DTW 실행 부분과 외부 메모리 인터페이스 부분으로 구성되어있다. CMOS 0.35um TLM 공정으로 설계된 이 칩은 4x4mm2의 면적에 126,938개의 게이트로 만들어져 있다. 그리고 전용 H/W의 동작 속도는 5MHz에서 60MHz까지 조정 가능하다. 5MHz 클록을 사용하는 경우, 50∼60 프레임 정도의 소어휘 단어 단위의 음성을 초당 100,000개까지 비교할 수 있는 능력이 있고, 60MHz의 클록을 사용하는 경우는 초당 1,200,000개의 단어를 비교할 수 있다. A speech recognition chip that can recognize a small vocabulary as a word-level has been designed. It is composed of EPD(Start and End-point detection) block, LPC block, DTW block and external memory interface block. It is made of 126,938 gates on 4x4mm2 area with a CMOS 0.35um TLM process. The speed of the chip varies from 5MHz to 60MHz because of its specific hardware designed for the purpose. It can compare 100,000 voices as a small vocabulary which has approximately 50∼60 frames at the clock of 5MHz and also up to 1,200,000 voices at the clock of 60MHz.

피-모스 트랜지스터에서 핫-캐리어에 의한 유효 이동도와 전계 효과 이동도의 분석

이용재,안점영 東義大學校 産業技術開發硏究所 1997 産業技術硏究誌 Vol.11 No.-

In this paper, the theoretical and empirical expressions most commonly used for modeling the variation of the low field mobility and field effect mobility of p-MOSFET's are discussed. It is shown that both approaches may be harmonized, and a new physical definition of the parameters of empirical models are presented. The mobility of hot-carrier-induced degradations by direct current stress has been characterized for p-MOSFET's with thin gate oxide. The effective mobility is derived from drain conductances, while the field-effect mobility is obtained from the transconductance of current-voltages characteristics. The empirical results are analyzed that the measurement data are identical at the point of maximum slop at near threshold voltage and the other part is different, that is, the field effect mobility is the slower than the effective mobility. It is found that the effect mobility and field effect mobility in p-MOSFET's are increased by increasing stress time and decreasing channel length.

회귀예측신경망과 확률신경망 결합에 의한 음절인식

어태경,김주성,안점영 東義大學校産業技術開發硏究所 1998 産業技術硏究誌 Vol.12 No.-

In this paper, we perform the recognition experiments of Korean 100 syllables based on RPNN. RPNN is a non-linear predictor according to the time variation of a speech without an alignment procedure. In the experiments, we study on the recognition rates, by increasing the prediction order and the number of hidden unit, and by appling the PNN as recognizer. In the method 1, when the prediction order is 3rd order and the hidden unit are 10 units, the recognition rates show better result than others. The recognition rates of the method 2 are improved than its results of the method 1, especially in the method 2, when the prediction order is 3rd order and the hidden units are 5 units, the recognition rates show better result than others.

音聲合成을 利用한 自動警報 裝置의 開發

배종갑,박재호,김정석,이종혁,안점영 東義大學校産業技術開發硏究所 1989 産業技術硏究誌 Vol.3 No.-

As society becomes more complex and nuclear families are increased by innovation of science and technology, many accidents like fire and robbery happen frequently leaving factory, shop and house empty. The necessity of a system that immediately alarms accidents taking place in the empty factory, shop and house, that enables to take necessary measures on the accidents to reduce the damages, is urgently raised. To meet such a need the automatic alarm device using speech synthesis is developed. According to the results of simulation verifying the validity of this system, the sensor of which radius and angle is 10[m]and 80˚respectively can detect the emergency and the system using this sensor can inform emergency signal to the reserved place very well.

출력층회귀 다층신경망의 숫자음 인식에 관한 연구

김영재,안점영 東義大學校 産業技術開發硏究所 1999 産業技術硏究誌 Vol.13 No.-

We compose the output layer recurrent multi-layered neural network that returns the output of output layer in the multi-layered perceptron with 4 layers to the lower hidden layer and study on the comparison of the spoken digit recognition performance according to the variation of the number of neuron in the upper and lower hidden layer, and the predictive order, and the learning rate and self-recurrent coefficient of the state layer. By the experimental results, when the number of neuron in the lower hidden layer is more than the number of neuron in the upper hidden layer or two layers have the same number of neuron, this network improves in it's recognition ability. The predictive order doesn't contribute to the improvement of the recognition rate. In the case of the learning rate and self-recurrent coefficient, the recognition rate is increased at 0.0001 and 0, respectively.