• MALSORI
• /
• no.60
• /
• pp.145-164
• /
• 2006

This paper proposes class histogram equalization (CHEQ) to compensate noisy acoustic features for robust speech recognition. CHEQ aims to compensate for the acoustic mismatch between training and test speech recognition environments as well as to reduce the limitations of the conventional histogram equalization (HEQ). In contrast to HEQ, CHEQ adopts multiple class-specific distribution functions for training and test environments and equalizes the features by using their class-specific training and test distributions. According to the class-information extraction methods, CHEQ is further classified into two forms such as hard-CHEQ based on vector quantization and soft-CHEQ using the Gaussian mixture model. Experiments on the Aurora 2 database confirmed the effectiveness of CHEQ by producing a relative word error reduction of 61.17% over the baseline met-cepstral features and that of 19.62% over the conventional HEQ.

• Journal of information and communication convergence engineering
• /
• v.19 no.3
• /
• pp.148-154
• /
• 2021

With the advent of context-aware computing, many attempts were made to understand emotions. Among these various attempts, Speech Emotion Recognition (SER) is a method of recognizing the speaker's emotions through speech information. The SER is successful in selecting distinctive 'features' and 'classifying' them in an appropriate way. In this paper, the performances of SER using neural network models (e.g., fully connected network (FCN), convolutional neural network (CNN)) with Mel-Frequency Cepstral Coefficients (MFCC) are examined in terms of the accuracy and distribution of emotion recognition. For Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) dataset, by tuning model parameters, a two-dimensional Convolutional Neural Network (2D-CNN) model with MFCC showed the best performance with an average accuracy of 88.54% for 5 emotions, anger, happiness, calm, fear, and sadness, of men and women. In addition, by examining the distribution of emotion recognition accuracies for neural network models, the 2D-CNN with MFCC can expect an overall accuracy of 75% or more.

• The Journal of the Acoustical Society of Korea
• /
• v.28 no.2E
• /
• pp.43-50
• /
• 2009

To get more natural synthetic speech generated by a Korean TTS (Text-To-Speech) system, we have to know all the possible prosodic rules in Korean spoken language. We should find out these rules from linguistic, phonetic information or from real speech. In general, all of these rules should be integrated into a prosody-generation algorithm in a TTS system. But this algorithm cannot cover up all the possible prosodic rules in a language and it is not perfect, so the naturalness of synthesized speech cannot be as good as we expect. ANNs (Artificial Neural Networks) can be trained to learn the prosodic rules in Korean spoken language. To train and test ANNs, we need to prepare the prosodic patterns of all the phonemic segments in a prosodic corpus. A prosodic corpus will include meaningful sentences to represent all the possible prosodic rules. Sentences in the corpus were made by picking up a series of words from the list of PB (phonetically Balanced) isolated words. These sentences in the corpus were read by speakers, recorded, and collected as a speech database. By analyzing recorded real speech, we can extract prosodic pattern about each phoneme, and assign them as target and test patterns for ANNs. ANNs can learn the prosody from natural speech and generate prosodic patterns of the central phonemic segment in phoneme strings as output response of ANNs when phoneme strings of a sentence are given to ANNs as input stimuli.

• Proceedings of the KSPS conference
• /
• 2006.05a
• /
• pp.37-40
• /
• 2006

In this paper, we propose a VQ codebook design of speech recognition feature parameters in order to improve the performance of a distributed speech recognition system. For the context-dependent HMMs, a VQ codebook should be correlated with phonetic distributions in the training data for HMMs. Thus, we focus on a selection method of training data based on phonetic distribution instead of using all the training data for an efficient VQ codebook design. From the speech recognition experiments using the Aurora 4 database, the distributed speech recognition system employing a VQ codebook designed by the proposed method reduced the word error rate (WER) by 10% when compared with that using a VQ codebook trained with the whole training data.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.37 no.7C
• /
• pp.538-543
• /
• 2012

The performance of the speech recognition engine is one of the most critical elements of the in-vehicle speech recognition interface. The objective of this paper is to develop an automated platform for running performance tests on the in-vehicle speech recognition engine. The developed platform comprise of main program, agent program, database management module, and statistical analysis module. A simulation environment for performance tests which mimics the real driving situations was constructed, and it was tested by applying pre-recorded driving noises and a speaker's voice as inputs. As a result, the validity of the results from the speech recognition tests was proved. The users will be able to perform the performance tests for the in-vehicle speech recognition engine effectively through the proposed platform.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.20 no.3
• /
• pp.471-476
• /
• 2016

This paper presents our study on speech recognition performance prediction. Our initial study shows that a combination of speech quality measures effectively improves correlation with Word Error Rate (WER) compared to each speech measure alone. In this paper we demonstrate a new combination of various types of speech quality measures shows more significantly improves correlation with WER compared to the speech measure combination of our initial study. In our study, SNR, PESQ, acoustic model score, and MFCC distance are used as the speech quality measures. This paper also presents our speech database verification system for speech recognition employing the speech measures. We develop a WER prediction system using Gaussian mixture model and the speech quality measures as a feature vector. The experimental results show the proposed system is highly effective at predicting WER in a low SNR condition of speech babble and car noise environments.

• The Journal of the Acoustical Society of Korea
• /
• v.14 no.4
• /
• pp.74-81
• /
• 1995

This paper describes the development of speech databases for the Korean language which were constructed at Communications Research Laboratory in KAIST. The procedure and environment to construct the speech database are presented in detail, and the phonetic and linguistic properties of the databases are presented. the databases were intended for use in designing and evaluating speech recognition algorithms. The databases consist of five different sets of speech contents : trade-related continuous speech with 3,000 words, variable-length connected digits, phoneme-balanced 75 isolated words, 500 isolated Korean provincial names, and Korean A-set words.

• The Journal of the Korea Contents Association
• /
• v.13 no.3
• /
• pp.1-8
• /
• 2013

This paper proposes a new feature extraction algorithm to improve children's speech recognition in Korean. The proposed feature extraction algorithm combines three methods. The first method is on the vocal tract length normalization to compensate acoustic features because the vocal tract length in children is shorter than in adults. The second method is to use the uniform bandwidth because children's voice is centered on high spectral regions. Finally, the proposed algorithm uses a smoothing filter for a robust speech recognizer in real environments. This paper shows the new feature extraction algorithm improves the children's speech recognition performance.

• Journal of the Korean Institute of Telematics and Electronics B
• /
• v.30B no.2
• /
• pp.51-58
• /
• 1993

This paper describes the development of a speaker-dependent isolated word recognizer as applied to voice dialing in a car noise environment. for this purpose, several methods to improve performance under such condition are evaluated using database collected in a small car moving at 100km/h The main features of the recognizer are as follow: The endpoint detection error can be reduced by using the magnitude of the signal which is inverse filtered by the AR model of the background noise, and it can be compensated by using variants of the DTW algorithm. To remove the noise, an autocorrelation subtraction method is used with the constraint that residual energy obtainable by linear predictive analysis should be positive. By using the noise rubust distance measure, distortion of the feature vector is minimized. The speech recognizer is implemented using the Motorola DSP56001(24-bit general purpose digital signal processor). The recognition database is composed of 50 Korean names spoken by 3 male speakers. The recognition error rate of the system is reduced to 4.3% using a single reference pattern for each word and 1.5% using 2 reference patterns for each word.

• Proceedings of the KIEE Conference
• /
• 2008.04a
• /
• pp.99-100
• /
• 2008

This paper proposes a Gaussian Mixture Model (GMM) - based speech emotion recognition methods using four feature parameters; 1) Fast Fourier Transform(FFT) spectral entropy, 2) delta FFT spectral entropy, 3) Mel-frequency Filter Bank (MFB) spectral entropy, and 4) delta MFB spectral entropy. In addition, we use four emotions in a speech database including anger, sadness, happiness, and neutrality. We perform speech emotion recognition experiments using each pre-defined emotion and gender. The experimental results show that the proposed emotion recognition using FFT spectral-based entropy and MFB spectral-based entropy performs better than existing emotion recognition based on GMM using energy, Zero Crossing Rate (ZCR), Linear Prediction Coefficient (LPC), and pitch parameters. In experimental Results, we attained a maximum recognition rate of 75.1% when we used MFB spectral entropy and delta MFB spectral entropy.