• The Journal of the Acoustical Society of Korea
• /
• v.25 no.8
• /
• pp.370-374
• /
• 2006

Mel-frequency cepstral coefficients are widely used as the feature for speech recognition. In FMCC extraction process. the spectrum. obtained by Fourier transform of input speech signal is divided by met-frequency bands, and each band energy is extracted for the each frequency band. The coefficients are extracted by the discrete cosine transform of the obtained band energy. In this Paper. we calculate the output energy for each bandpass filter by taking the weighting function when applying met-frequency scaled bandpass filter. The weighting function is Gaussian distributed function whose center is at the formant frequency In the experiments, we can see the comparative performance with the standard MFCC in clean condition. and the better Performance in worse condition by the method proposed here.

• The Journal of the Acoustical Society of Korea
• /
• v.21 no.3
• /
• pp.304-309
• /
• 2002

We propose the computation reduction method of real root method that is mainly used in the CELP (Code Excited Linear Prediction) vocoder. The real root method is that if polynomial equations have the real roots, we are able to find those and transform them into LSP. However, this method takes much time to compute, because the root searching is processed sequentially in frequency region. In this paper, to reduce the computation time of real root, we compare the real root method with two methods. In first method, we use the mal scale of searching frequency region that is linear below 1 kHz and logarithmic above. In second method, The searching frequency region and searching interval are ordered by each coefficient's distribution. In order to compare real root method with proposed methods, we measured the following two. First, we compared the position of transformed LSP (Line Spectrum Pairs) parameters in the proposed methods with these of real root method. Second, we measured how long computation time is reduced. The experimental results of both methods that the searching time was reduced by about 47% in average without the change of LSP parameters.

• Journal of the Institute of Convergence Signal Processing
• /
• v.9 no.2
• /
• pp.121-128
• /
• 2008

Accurate voice activity detection have a great impact on performance of speech applications including speech recognition, speech coding, and speech communication. In this paper, we propose methods for voice activity detection that can adapt to various car noise situations during driving. Existing voice activity detection used various method such as time energy, frequency energy, zero crossing rate, and spectral entropy that have a weak point of rapid. decline performance in noisy environments. In this paper, the approach is based on existing spectral entropy for VAD that we propose voice activity detection method using MFB(Met-frequency filter banks) spectral entropy, gradient FFT(Fast Fourier Transform) spectral entropy. and gradient MFB spectral entropy. FFT multiplied by Mel-scale is MFB and Mel-scale is non linear scale when human sound perception reflects characteristic of speech. Proposed MFB spectral entropy method clearly improve the ability to discriminate between speech and non-speech for various in noisy car environments that achieves 93.21% accuracy as a result of experiments. Compared to the spectral entropy method, the proposed voice activity detection gives an average improvement in the correct detection rate of more than 3.2%.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.6
• /
• pp.600-605
• /
• 2020

In this paper, we propose a sound event detection method using a multi-channel multi-scale neural networks for sound sensing home monitoring for the hearing impaired. In the proposed system, two channels with high signal quality are selected from several wireless microphone sensors in home. The three features (time difference of arrival, pitch range, and outputs obtained by applying multi-scale convolutional neural network to log mel spectrogram) extracted from the sensor signals are applied to a classifier based on a bidirectional gated recurrent neural network to further improve the performance of sound event detection. The detected sound event result is converted into text along with the sensor position of the selected channel and provided to the hearing impaired. The experimental results show that the sound event detection method of the proposed system is superior to the existing method and can effectively deliver sound information to the hearing impaired.

• Proceedings of the Acoustical Society of Korea Conference
• /
• 1993.06a
• /
• pp.113-116
• /
• 1993

본 논문에서는 변형된 wavelet 변환을 통해 추출한 특징벡터를 이용하여 한국어 숫자음을 대상으로 한 음성인식기를 구현하였다. wavelet 변환은 시간 및 주파수 영역에 대해 다중해상도(multiresolution)를 가지는 신호분석법이다. 본 연구에서는 계산량의 감소와 넓은 주파수 대역을 분석하기 위해, mother wavelet의 형태를 분석 주파수 대역에 따라 변화시키는 방법을 제안하였다. 기존의 wavelet 변환으로 실험한 결과 86.5%의 인식율을 얻었고, 변형된 wavelet 변환의 경우 96%의 인식율을 얻었으며 계산량이 감소하였다. 이와 함께 음성인식에서 널리 사용되는 특징 파라미터인 멜켑스트럼과 FFT 멜스케일 필터 대역(mel scale filter bank)과 비교 실험한 결과 인식율의 향상을 보였다. 이는 제안한 방법이 고주파 대역의 세밀한 시간 해상도와 저주파 대역의 세밀한 주파수 해상도를 지니는데 기인하는 것으로 판단된다.

• Proceedings of the KAIS Fall Conference
• /
• 2010.05a
• /
• pp.249-252
• /
• 2010

LPC 계수를 LSP 변환하는 방법에는 복소근, 실근, 비율 필터, 체비셰프 급수, 적응적 순차형 최소제곱 평균 방법(adaptive sequential LMS) 등이 있다. 이 방법들 중 음성 부호화기에서 주로 사용하는 실근 방법은 근을 구하기 위해 주파수 영역을 순차적으로 검색하기 때문에 계산시간이 많이 소요되는 단점을 갖는다. 본 논문에서는 LPC에서 LSP로 변환하는 4가지 고속 알고리즘을 제안한다. 첫 번째 방식에서는 검색간격에 멜 스케일을 적용하였고, 두 번째는 홀수번째 LSP 파라미터의 분포도를 이용하여 검색순서를 조정한 방법이다. 세 번째 방식과 네 번째 방식에서는 각각, 모음 특성, LSP 분포특성과 해상도를 이용하여 계산시간을 단축하였다. LSP 변환시간은 4가지 방법 모두 35~50% 단축되었다. 또한 실험결과에서는 각 알고리즘의 고유한 특성에 대하여 분석한다.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2009.10a
• /
• pp.705-708
• /
• 2009

This research describes a neural network-based phoneme segmenter for recognizing spontaneous speech. The input of the phoneme segmenter for spontaneous speech is 16th order mel-scaled FFT, normalized frame energy, ratio of energy among 0~3[KHz] band and more than 3[KHz] band. All the features are differences of two consecutive 10 [msec] frame. The main body of the segmenter is single-hidden layer MLP(Multi-Layer Perceptron) with 72 inputs, 20 hidden nodes, and one output node. The segmentation accuracy is 78% with 7.8% insertion.

• The Journal of the Acoustical Society of Korea
• /
• v.22 no.4
• /
• pp.299-307
• /
• 2003

There have been many studies on speaker normalization which aims to minimize the effects of speaker's vocal tract length on the recognition performance of the speaker independent speech recognition system. In this paper, we propose a simple vector quantizer based linear warping speaker normalization method based on the observation that the vector quantizer can be successfully used for speaker verification. For this purpose, we firstly generate an optimal codebook which will be used as the basis of the speaker normalization, and then the warping factor of the unknown speaker will be extracted by comparing the feature vectors and the codebook. Finally, the extracted warping factor is used to linearly warp the Mel scale filter bank adopted in the course of MFCC calculation. To test the performance of the proposed method, a series of recognition experiments are conducted on discrete HMM with thirteen mono-syllabic Korean number utterances. The results showed that about 29% of word error rate can be reduced, and that the proposed warping factor extraction method is useful due to its simplicity compared to other line search warping methods.

• Journal of Digital Contents Society
• /
• v.18 no.1
• /
• pp.183-190
• /
• 2017

This paper proposes an audio event detection method using Deep Neural Networks (DNN). The proposed method applies Feed Forward Neural Network (FFNN) to generate output probabilities of twenty audio events for each frame. Mel scale filter bank (FBANK) features are extracted from each frame, and its five consecutive frames are combined as one vector which is the input feature of the FFNN. The output layer of FFNN produces audio event probabilities for each input feature vector. More than five consecutive frames of which event probability exceeds threshold are detected as an audio event. An audio event continues until the event is detected within one second. The proposed method achieves as 71.8% accuracy for 20 classes of the UrbanSound8K and the BBC Sound FX dataset.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.8 no.3
• /
• pp.563-568
• /
• 2007

We propose the computation reduction method of real root method that is used in the EVRC(Enhanced Variable Rate Codec) system. The real root method is that if polynomial equations have the real roots, we are able to find those and transform them into LSP. However, this method takes much time to compute, because the root searching is processed sequentially in frequency region. But, the important characteristic of LSP is that most of coefficients are occurred in specific frequency region. So, to reduce the computation time of real root, we used the met scale that is linear below 1kHz and logarithmic above. In order to compare real root method with proposed method, we measured the following two. First, we compared the position of transformed LSP(Line Spectrum Pairs) parameters in the proposed method with these of real root method. Second, we measured how long computation time is reduced. The experimental result is that the searching time was reduced by about 48% in average without the change of LSP parameters.