• Proceedings of the Acoustical Society of Korea Conference
• /
• 1994.06c
• /
• pp.403-406
• /
• 1994

음성은 화자의 상태 및 주변 환경에 따라 그 특징이 다양하게 변화한다. 본 논문에서는 음성신호의 특징 파라미터로 널리 쓰이고 있는 mel-cepstrum에 대해, 단어내에서의 변화를 정규화함으로써 인식성능을 향상시키고자 하였다. mel-cepstrum이란 단어 전체에 대한 mel-cepstrum의 평균 값으로 normalize 시킨 것이다. 한국어 숫자음에 대한 인식 실험결과, 본 논문에서 제안한 정규화된 mel-cepstrum이 정규화되지 않은 mel-cepstrum에 비해 우수한 인식 성능을 나타내었다. 또한 잡음 환경하에서 비교 실험한 결과에서도 상대적으로 우수한 인식률을 보였다.

• Speech Sciences
• /
• v.2
• /
• pp.57-66
• /
• 1997

This paper studies the effects of the method, CMS(Cepstral Mean Subtraction), (which compensates for some of the speech distortion. caused by telephone channels), on the performance of the text-dependent speaker identification system. This system is based on the VQ(Vector Quantization) and HMM(Hidden Markov Model) method and chooses the LPC-Cepstrum and Mel-Cepstrum as the feature vectors extracted from the speech data transmitted through telephone channels. Accordingly, we can compare the correct recognition rates of the speaker identification system between the use of LPC-Cepstrum and Mel-Cepstrum. Finally, from the experiment results table, it is found that the Mel-Cepstrum parameter is proven to be superior to the LPC-Cepstrum and that recognition performance improves by about 10% when compensating for telephone channel using the CMS.

• MALSORI
• /
• no.64
• /
• pp.89-103
• /
• 2007

In this paper, we introduce mel-cepstrum modulation energy (MCME) for a feature to discriminate speech and music data. MCME is a mel-cepstrum domain extension of modulation energy (ME). MCME is extracted on the time trajectory of Mel-frequency cepstral coefficients, while ME is based on the spectrum. As cepstral coefficients are mutually uncorrelated, we expect the MCME to perform better than the ME. To find out the best modulation frequency for MCME, we perform experiments with 4 Hz to 20 Hz modulation frequency. To show effectiveness of the proposed feature, MCME, we compare the discrimination accuracy with the results obtained from the ME and the cepstral flux.

• ETRI Journal
• /
• v.18 no.1
• /
• pp.29-41
• /
• 1996

This paper describes a method to select a suitable feature for speech recognition using information theoretic measure. Conventional speech recognition systems heuristically choose a portion of frequency components, cepstrum, mel-cepstrum, energy, and their time differences of speech waveforms as their speech features. However, these systems never have good performance if the selected features are not suitable for speech recognition. Since the recognition rate is the only performance measure of speech recognition system, it is hard to judge how suitable the selected feature is. To solve this problem, it is essential to analyze the feature itself, and measure how good the feature itself is. Good speech features should contain all of the class-related information and as small amount of the class-irrelevant variation as possible. In this paper, we suggest a method to measure the class-related information and the amount of the class-irrelevant variation based on the Shannon's information theory. Using this method, we compare the mel-scaled FFT, cepstrum, mel-cepstrum, and wavelet features of the TIMIT speech data. The result shows that, among these features, the mel-scaled FFT is the best feature for speech recognition based on the proposed measure.

• Proceedings of the IEEK Conference
• /
• 2001.09a
• /
• pp.833-836
• /
• 2001

본 논문에서는 한국어 음성인식 시스템의 성능 향상을 위해 청각 주파수 분해능을 가진 MEL-LPC Cepstrum을 음소단위의 HMM(Hidden Markov Model)을 기반으로 하는 인식 시스템에 적용하여 그 결과를 비교 검토하였다. 선형예측(LP) 분석 후에 후처리로서 주파수를 왜곡시킨 LPC-MEL 분석이 계산량이 적고 효과적이라 일반적으로 많이 사용되고 있으나 주파수 분해능은 많이 개선되지 않는다. 따라서 본 논문에서는 주파수 분해능을 개선하기 위해, 원 음성신호로부터 직접적으로 멜주파수로 왜곡시킨 후 선형 예측 분석을 수행하는 MEL-LPC 분석방법을 이용한 음소기반의 화자 독립 음성인식 시스템을 구성하여 기존의 LPC-MEL 분석방법과 비교실험을 통하여 MEL-LPC 분석방법의 유효성을 검토하였다. 실험에 사용한 음성 데이터베이스는 음소 및 단어 인식실험에서는 ETRI 445단어 DB, 연속 숫자음인식 실험에서는 KLE 4연속 숫자음 DB를 사용하였다. 화자 독립 음소인식 실험의 경우, 묵음을 제외한 47개의 유사 음소에 대하여 4상태 3출력의 Left-to-Right 모델을이용하였다. 단어 및 연속 숫자음 인식 실험의 경우, 유한상태 네트워크에 의한 OPDP법을 이용하였다. 화자 독립 음소, 단어 및 4연속 숫자음 인식 실험결과, 기존의 LPC-MEL Cepstrum을 사용한 경우보다 MEL-LPC Cepstum을 사용한 경우가 더 높은 인식률을 나타내어 한국어 음성인식 시스템에서 MEL-LPC 분석방법의 유효성을 확인할 수 있었다.

• The Journal of the Acoustical Society of Korea
• /
• v.18 no.1
• /
• pp.78-85
• /
• 1999

In this paper we propose a method to improve the performance of the mel-cepstrum that is widely used in speech recognition. Typically, the met-cepstrum is obtained by critical band filters that have fixed center spacing and bandwidth. However different filter characteristics produce a different mel-cepstrum, resulting in a different performance. In this paper we analyze triangular-shaped and rectangular-shaped filters. By changing the characteristics of filters such as center frequency and bandwidth, we analyze the performance of the met-cepstrum. Then utilizing the simplex method, we propose a method to optimize the critical band filters. Using the dynamic time warping, we performed speaker independent recognition experiments with Korean digit words pronounced by 10 males and 10 females. Experiments show that the rectangular-shaped filters show good performance and the mel-cepstrum obtained by the optimized filters shows better performance than filters that have fixed center spacing and bandwidth.

• Proceedings of the Acoustical Society of Korea Conference
• /
• 1994.06c
• /
• pp.157-160
• /
• 1994

본 논문에서는 한국어 숫자음 인식에 대한 이산분포 HMM과 연속분포 HMM의 인식 성능을 비교하였다. 일반적으로 연속분포 HMM은 많은 계산량이 필요하고, 학습시 초기값이 매우 민감하다는 단점이 있지만, 이산분포 HMM의 VQ로 인한 왜곡을 제거함으로써 인식률을 향상시킬 수 있다. 여기서는 성능비교를 위해서 mel-cepstrum의 분석차수, 이산분포 HMM의 codebook 크기, 연속분포 HMM의 miture 개수등에 따른 인식성능을 비교하였다. 실험 결과 이산분포 HMM에서는 mel-cepstrum 벡터가 14차이고, codebook 크기가 64일 때 가장 좋은 성능을 나타냈으며, 연속부포 HMM에서는 mel-cepstrum 벡터가 16차이고 miture가 3개일 때 가장 좋은 결과를 얻을 수 있었다. 특히 학습 데이터의 양이 적은 경우에는 연속분포 HMM이 이산분포 HMM보다 더 좋은 인식률을 나타내었다.

• The Journal of the Acoustical Society of Korea
• /
• v.13 no.1E
• /
• pp.71-81
• /
• 1994

In this paper, we propose a pseudo-cepstral representation of line spectrum pair(LSP) frequencies and evaluate speech recognition performance with cepstral lift using the pseudo-cepstrum. The pseudo-cepstrum corresponding to LSP frequencies is derived by approxmating the relationship between LPC-cepstrum and LSP frequencies. Three cepstral liftering procedures are applied to the pseudo-cepstrum to improve the performance of speech recognition. They are the root-power-sums ligter, the general exponential lifter, and the bandpass lifter. Then, the liftered psedudo-cepstra are warped into a mel-frequency scale to obtain feature vectors for speech recognition. Among the three lifters, the general exponential lifter results in the best performance on speech recognition. When we use the proposed pseudo-cepstra feature vectors for recognizing noisy speech, the signal-to-noise ratio (SNR) improvement of about 5~10dB LSP is obtained.

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.2
• /
• pp.10-15
• /
• 2001

In this paper, we describe the comparison between the combination of features using a speech and music discrimination, which is classifying between speech and music on audio signals. Audio signals are classified into 3classes (speech, music, speech and music) and 2classes (speech, music). Experiments carried out on three types of feature, Mel-cepstrum, energy, zero-crossings, and try to find a best combination between features to speech and music discrimination. We using a Gaussian Mixture Model (GMM) for discrimination algorithm and combine different features into a single vector prior to modeling the data with a GMM. In 3classes, the best result is achieved using Mel-cepstrum, energy and zero-crossings in a single feature vector (speech: 95.1％, music: 61.9％, speech & music: 55.5％). In 2classes, the best result is achieved using Mel-cepstrum, energy and Mel-cepstrum, energy, zero-crossings in a single feature vector (speech: 98.9％, music: 100％).

• Proceedings of the Acoustical Society of Korea Conference
• /
• 1995.06a
• /
• pp.225-227
• /
• 1995

This paper compared waveform, cepstrum, and spline wavelet features with nonlinear discriminant analysis. This measure shows efficiency of speech parametrization better than old linear separability criteria and can be used to measure the efficiency of each layer of certain system. Spline wavelet transform has larger gap among classes and cepstrum is clustered better than the spline wavelet feature. Both features do not have good property for classification and we will compare Gabor wavelet transform, Mel cepstrum, delta cepstrum, etc.