• Phonetics and Speech Sciences
• /
• v.6 no.3
• /
• pp.155-164
• /
• 2014

This paper proposes a new method to determine the recognition units for large vocabulary continuous speech recognition (LVCSR) in Korean by applying unsupervised segmentation and merging. In the proposed method, a text sentence is segmented into morphemes and position information is added to morphemes. Then submorpheme units are obtained by splitting the morpheme units through the maximization of posterior probability terms. The posterior probability terms are computed from the morpheme frequency distribution, the morpheme length distribution, and the morpheme frequency-of-frequency distribution. Finally, the recognition units are obtained by sequentially merging the submorpheme pair with the highest frequency. Computer experiments are conducted using a Korean LVCSR with a 100k word vocabulary and a trigram language model obtained by a 300 million eojeol (word phrase) corpus. The proposed method is shown to reduce the out-of-vocabulary rate to 1.8% and reduce the syllable error rate relatively by 14.0%.

• Proceedings of the IEEK Conference
• /
• 2000.07a
• /
• pp.257-259
• /
• 2000

In this paper, a word-unit recognition is performed to implement a speech recognition system over the web, using JAVA Applet and continuous distributed HMM. The system based on Client/server model is designed. A client computer processes speech with Applet, and then transmits feature parameters to the server computer though the Internet. The speech recognition system in the server computer transmits the result applied by the forward algorithm to the client computer and the result is displayed in the client computer by text.

• Proceedings of the IEEK Conference
• /
• 2006.06a
• /
• pp.259-260
• /
• 2006

Since audio speech recognition is affected by noise in real environment, visual speech recognition is used to support speech recognition. For the visual speech recognition, this paper suggests the extraction of lip-feature using two types of image segmentation and reduced ASM. Input images are transformed to YCbCr based images and lips are segmented using the contrast of Y/Cb/Cr between lip and face. Subsequently, lip-shape model trained by PCA is placed on segmented lip region and then lip features are extracted using ASM.

• The Journal of the Acoustical Society of Korea
• /
• v.38 no.5
• /
• pp.601-606
• /
• 2019

This paper is a study on speech recognition in the Korean using grapheme unit (Cho-sumg [onset], Jung-sung [nucleus], Jong-sung [coda]). Here we make ASR (Automatic speech recognition) system without G2P (Grapheme to Phoneme) process and show that Deep learning based ASR systems can learn Korean pronunciation rules without G2P process. The proposed model is shown to reduce the word error rate in the presence of sufficient training data.

• The Journal of the Acoustical Society of Korea
• /
• v.18 no.3
• /
• pp.62-67
• /
• 1999

In this paper, we investigates a reliable model of the Predictive Recurrent Neural Network for the speech recognition. Predictive Neural Networks are modeled by syllable units. For the given input syllable, then a model which gives the minimum prediction error is taken as the recognition result. The Predictive Neural Network which has the structure of recurrent network was composed to give the dynamic feature of the speech pattern into the network. We have compared with the recognition ability of the Recurrent Network proposed by Elman and Jordan. ETRI's SAMDORI has been used for the speech DB. In order to find a reliable model of neural networks, the changes of two recognition rates were compared one another in conditions of: (1) changing prediction order and the number of hidden units: and (2) accumulating previous values with self-loop coefficient in its context. The result shows that the optimum prediction order, the number of hidden units, and self-loop coefficient have differently responded according to the structure of neural network used. However, in general, the Jordan's recurrent network shows relatively higher recognition rate than Elman's. The effects of recognition rate on the self-loop coefficient were variable according to the structures of neural network and their values.

• Proceedings of the IEEK Conference
• /
• 2000.06d
• /
• pp.127-130
• /
• 2000

In this paper, we explore the possibility to use wavelet decomposition based on modified octave structured 5-level filter banks as a set of features for speech recognition. The HMM (Hidden Markov Model) is used as a recognizer 〔l〕. We compared the performance of the wavelet decomposition with the mel-cepstrum and LPC cepstrum. Experimental results show favorable results.

• Proceedings of the KSPS conference
• /
• 2003.05a
• /
• pp.83-86
• /
• 2003

This paper presents our style-based language model adaptation for Korean conversational speech recognition. Korean conversational speech is observed various characteristics of content and style such as filled pauses, word omission, and contraction as compared with the written text corpora. For style-based language model adaptation, we report two approaches. Our approaches focus on improving the estimation of domain-dependent n-gram models by relevance weighting out-of-domain text data, where style is represented by n-gram based tf*idf similarity. In addition to relevance weighting, we use disfluencies as predictor to the neighboring words. The best result reduces 6.5％ word error rate absolutely and shows that n-gram based relevance weighting reflects style difference greatly and disfluencies are good predictor.

• Proceedings of the KIEE Conference
• /
• 1990.07a
• /
• pp.535-538
• /
• 1990

The main problem in speech recognition is the enormous variability in acoustic signals due to complex but predictable contextual effects. Especially in plosive consonants it is very difficult to find invariant cue due to various contextual effects, but humans use these contextual effects as helpful information in plosive consonant recognition. In this paper we experimented on three artificial neural net models for the recognition of plosive consonants. Neural Net Model I used "Multi-layer Perceptron ". Model II used a variation of the "Self-organizing Feature Map Model". And Model III used "Interactive and Competitive Model" to experiment contextual effects. The recognition experiment was performed on 9 Korean plosive consonants. We used VCV speech chains for the experiment on contextual effects. The speech chain consists of Korean plosive consonants /g, d, b, K, T, P, k, t, p/ (/ㄱ, ㄷ, ㅂ, ㄲ, ㄸ, ㅃ, ㅋ, ㅌ, ㅍ/) and eight Korean monothongs. The inputs to Neural Net Models were several temporal cues - duration of the silence, transition and vot -, and the extent of the VC formant transitions to the presence of voicing energy during closure, burst intensity, presence of asperation, amount of low frequency energy present at voicing onset, and CV formant transition extent from the acoustic signals. Model I showed about 55 - 67 %, Model II showed about 60%, and Model III showed about 67% recognition rate.

• Journal of Convergence for Information Technology
• /
• v.11 no.6
• /
• pp.33-39
• /
• 2021

Speech recognition technology is being combined with deep learning and is developing at a rapid pace. In particular, voice recognition services are connected to various devices such as artificial intelligence speakers, vehicle voice recognition, and smartphones, and voice recognition technology is being used in various places, not in specific areas of the industry. In this situation, research to meet high expectations for the technology is also being actively conducted. Among them, in the field of natural language processing (NLP), there is a need for research in the field of removing ambient noise or unnecessary voice signals that have a great influence on the speech recognition recognition rate. Many domestic and foreign companies are already using the latest AI technology for such research. Among them, research using a convolutional neural network algorithm (CNN) is being actively conducted. The purpose of this study is to determine the non-voice section from the user's speech section through the convolutional neural network. It collects the voice files (wav) of 5 speakers to generate learning data, and utilizes the convolutional neural network to determine the speech section and the non-voice section. A classification model for discriminating speech sections was created. Afterwards, an experiment was conducted to detect the non-speech section through the generated model, and as a result, an accuracy of 94% was obtained.

• The Journal of the Acoustical Society of Korea
• /
• v.18 no.5
• /
• pp.78-82
• /
• 1999

This paper is a study on weighted cepstrum used broadly for robust speech recognition. Especially, we propose the weighted function of asymmetric glottal pulse shape. which is used for weighted cepstrum extracted by PLP(Perceptual Linear Predictive) based on auditory model. Also, we analyze this glottal weighted cepstrum from the glottal pulse of glottal model in connection with the cepstrum. And we obtain speech features analyzed by both the glottal model and the auditory model. The isolated-word recognition rate is adopted for the test of proposed method in the car moise and street environment. And the performance of glottal weighted cepstrum is compared with both that of weighted cepstrum extracted by LP(Linear Prediction) and that of weighted cepstrum extracted by PLP. The result of computer simulation shows that recognition rate of the proposed glottal weighted cepstrum is better than those of other weighted cepstrums.