• MALSORI
• /
• no.48
• /
• pp.141-153
• /
• 2003

Researches into robust recognition in noise environments, especially in car environments, are being carried out actively in speech community. In this paper we will report on three types of corpora that SiTEC (Speech Information TEchnology & industry promotion Center) has created for research into speech recognition in car noise environments. The first is the recordings of 900 Korean native speakers, distributed according to gender, age, and region, who uttered application words in car environments. The second is the collections of mixed noise in 3 car types by model while setting up various noise patterns which can be obtained with the car engine on or off, at different driving speed, and in different road conditions with windows open or closed. The third is the recordings of simulated speech by HATS (Head and Torso Simulator) in car environments with the internal and external noise factors added. These three types of recordings were all made through synchronized 8 channel microphones that are fixed in a car. The creation and applications of these corpora will be reported on in detail.

• Journal of Internet Computing and Services
• /
• v.24 no.2
• /
• pp.11-18
• /
• 2023

Recently, in the field of Speech Emotion Recognition (SER), many studies have been conducted to improve accuracy using voice features and modeling. In addition to modeling studies to improve the accuracy of existing voice emotion recognition, various studies using voice features are being conducted. This paper, voice files are separated by time interval in a time series method, focusing on the fact that voice emotions are related to time flow. After voice file separation, we propose a model for classifying emotions of speech data by extracting speech features Mel, Chroma, zero-crossing rate (ZCR), root mean square (RMS), and mel-frequency cepstrum coefficients (MFCC) and applying them to a recurrent neural network model used for sequential data processing. As proposed method, voice features were extracted from all files using 'librosa' library and applied to neural network models. The experimental method compared and analyzed the performance of models of recurrent neural network (RNN), long short-term memory (LSTM) and gated recurrent unit (GRU) using the Interactive emotional dyadic motion capture Interactive Emotional Dyadic Motion Capture (IEMOCAP) english dataset.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.25 no.3
• /
• pp.272-278
• /
• 2015

This paper studied various parameter extraction methods using pitch information of speech for the development of the emotion recognition system. For this purpose, pitch parameters were extracted from korean speech database containing various emotions using stochastical information and numerical analysis techniques. GMM based emotion recognition system were used to compare the performance of pitch parameters. Sequential feature selection method were used to select the parameters showing the best emotion recognition performance. Experimental results of recognizing four emotions showed 63.5% recognition rate using the combination of 15 parameters out of 56 pitch parameters. Experimental results of detecting the presence of emotion showed 80.3% recognition rate using the combination of 14 parameters.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.21 no.12
• /
• pp.2291-2297
• /
• 2017

This paper describe to extract speech measure algorithm for evaluating a speech database, and presents generating method of a speech quality measure using DNN(Deep Neural Network). In our previous study, to produce an effective speech quality measure, we propose a combination of various speech measures which are highly correlated with WER(Word Error Rate). The new combination of various types of speech quality measures in this study is more effective to predict the speech recognition performance compared to each speech measure alone. In this paper, we describe the method of extracting measure using DNN, and we change one of the combined measure from GMM(Gaussican Mixture Model) score used in the previous study to DNN score. The combination with DNN score shows a higher correlation with WER compared to the combination with GMM score.

• The Journal of the Acoustical Society of Korea
• /
• v.16 no.1E
• /
• pp.15-23
• /
• 1997

Zero-Crossings with Peak Amplitudes(ZCPA) model motivated by human auditory periphery was proposed to extract reliable features speech signals even in noisy environments for robust speech recognition. In this paper, some practical considerations for digital hardware implementations of the ZCPA model are addressed and evaluated for recognition of speech corrupted by several real world noises as well as white Gaussian noise. Infinite impulse response(IIR) filters which constitute the cochliar filterbank of the ZCPA are replaced by hamming bandpass filters of which frequency responses are less similar to biological neural tuning curves. Experimental results demonstrate that the detailed frequency response of the cochlear filters are not critical to performance. Also, the sensitivity of the model output to the variations in microphone gain is investigated, and results in good reliability of the ZCPA model.

• Phonetics and Speech Sciences
• /
• v.12 no.2
• /
• pp.29-37
• /
• 2020

Recently, the neural network-based deep learning algorithm has dramatically improved performance compared to the classical Gaussian mixture model based hidden Markov model (GMM-HMM) automatic speech recognition (ASR) system. In addition, researches on end-to-end (E2E) speech recognition systems integrating language modeling and decoding processes have been actively conducted to better utilize the advantages of deep learning techniques. In general, E2E ASR systems consist of multiple layers of encoder-decoder structure with attention. Therefore, E2E ASR systems require data with a large amount of speech-text paired data in order to achieve good performance. Obtaining speech-text paired data requires a lot of human labor and time, and is a high barrier to building E2E ASR system. Therefore, there are previous studies that improve the performance of E2E ASR system using relatively small amount of speech-text paired data, but most studies have been conducted by using only speech-only data or text-only data. In this study, we proposed a semi-supervised training method that enables E2E ASR system to perform well in corpus in different domains by using both speech or text only data. The proposed method works effectively by adapting to different domains, showing good performance in the target domain and not degrading much in the source domain.

• ETRI Journal
• /
• v.44 no.3
• /
• pp.476-490
• /
• 2022

With the recent advances in technology, automatic speech recognition (ASR) has been widely used in real-world applications. The efficiency of converting large amounts of speech into text accurately with limited resources has become more vital than ever. In this study, we propose a method to rapidly recognize a large speech database via a transformer-based end-to-end model. Transformers have improved the state-of-the-art performance in many fields. However, they are not easy to use for long sequences. In this study, various techniques to accelerate the recognition of real-world speeches are proposed and tested, including decoding via multiple-utterance-batched beam search, detecting end of speech based on a connectionist temporal classification (CTC), restricting the CTC-prefix score, and splitting long speeches into short segments. Experiments are conducted with the Librispeech dataset and the real-world Korean ASR tasks to verify the proposed methods. From the experiments, the proposed system can convert 8 h of speeches spoken at real-world meetings into text in less than 3 min with a 10.73% character error rate, which is 27.1% relatively lower than that of conventional systems.

• The Journal of the Acoustical Society of Korea
• /
• v.13 no.1
• /
• pp.14-23
• /
• 1994

In this paper, speech units appropriate for recognition of Korean language have been studied. For better speech recognition, co-articulatory effects within an utterance should be considered in the selection of a recognition unit. One way to model such effects is to use larger units of speech. It has been found that diphone is a good recognition unit because it can model transitional legions explicitly. When diphone is used, stationary phoneme models may be inserted between diphones. Computer simulation for isolated word recognition was done with 7 word database spoken by seven male speakers. Best performance was obtained when transition regions between phonemes were modeled by two-state HMM's and stationary phoneme regions by one-state HMM's excluding /b/, /d/, and /g/. By merging rarely occurring diphone units, the recognition rate was increased from $93.98\%$ to $96.29\%$. In addition, a local interpolation technique was used to smooth a poorly-modeled HMM with a well-trained HMM. With this technique we could get the recognition rate of $97.22\%$ after merging some diphone units.

• Proceedings of the KSPS conference
• /
• 2002.11a
• /
• pp.197-200
• /
• 2002

In this paper, we propose a phoneme duration modeling in a speech recognition system based on disicion tree state tying. We assume that phone duration has a Gamma distribution. In a training mode, we model mean and variance of each state duration in context-independent phone model based on decision tree state tying. In a recognition mode, we get mean and variance of each context-dependent phone duration form state duration information obtaind during training mode. We make a comparative study of the proposed meth with conventinal methods. Our method results in good performance compared with conventional methods.

• Journal of the Institute of Convergence Signal Processing
• /
• v.4 no.4
• /
• pp.30-39
• /
• 2003

In this paper, we carried out the performance evaluation of HM-Net(Hidden Markov Network) speech recognition system for Korean speech databases. We adopted to construct acoustic models using the HM-Nets modified by HMMs(Hidden Markov Models), which are widely used as the statistical modeling methods. HM-Nets are carried out the state splitting for contextual and temporal domain by PDT-SSS(Phonetic Decision Tree-based Successive State Splitting) algorithm, which is modified the original SSS algorithm. Especially it adopted the phonetic decision tree to effectively express the context information not appear in training speech data on contextual domain state splitting. In case of temporal domain state splitting, to effectively represent information of each phoneme maintenance in the state splitting is carried out, and then the optimal model network of triphone types are constructed by in the parameter. Speech recognition was performed using the one-pass Viterbi beam search algorithm with phone-pair/word-pair grammar for phoneme/word recognition, respectively and using the multi-pass search algorithm with n-gram language models for sentence recognition. The tree-structured lexicon was used in order to decrease the number of nodes by sharing the same prefixes among words. In this paper, the performance evaluation of HM-Net speech recognition system is carried out for various recognition conditions. Through the experiments, we verified that it has very superior recognition performance compared with the previous introduced recognition system.