• The Journal of the Acoustical Society of Korea
• /
• v.21 no.4
• /
• pp.407-414
• /
• 2002

In this paper, we present a method that automatically generates concatenate morpheme based language models to improve the performance of Korean large vocabulary continuous speech recognition. The focus was brought into improvement against recognition errors of monosyllable morphemes that occupy 54% of the training text corpus and more frequently mis-recognized. Knowledge-based method using POS patterns has disadvantages such as the difficulty in making rules and producing many low frequency concatenate morphemes. Proposed method automatically selects morpheme-pairs from training text data based on measures such as frequency, mutual information, and unigram log likelihood. Experiment was performed using 7M-morpheme text corpus and 20K-morpheme lexicon. The frequency measure with constraint on the number of morphemes used for concatenation produces the best result of reducing monosyllables from 54% to 30%, bigram perplexity from 117.9 to 97.3. and MER from 21.3% to 17.6%.

• The Journal of the Acoustical Society of Korea
• /
• v.22 no.5
• /
• pp.360-366
• /
• 2003

Although multilayer perceptrons (MLPs) present several advantages against other pattern recognition methods, MLP-based speaker verification systems suffer from slow enrollment speed caused by many background speakers to achieve a low verification error. To solve this problem, the quantitative discriminative cohort speakers (QnDCS) method, by introducing the cohort speakers method into the systems, reduced the number of background speakers required to enroll speakers. Although the QnDCS achieved the goal to some extent, the improvement rate for the enrolling speed was still unsatisfactory. To improve the enrolling speed, this paper proposes the qualitative DCS (QlDCS) by introducing a qualitative criterion to select less background speakers. An experiment for both methods is conducted to use the speaker verification system based on MLPs and continuants, and speech database. The results of the experiment show that the proposed QlDCS method enrolls speakers in two times shorter time than the QnDCS does over the online error backpropagation(EBP) method.

• Journal of Digital Convergence
• /
• v.13 no.10
• /
• pp.353-358
• /
• 2015

The Vocabulary Recognition System made by recognizing the standard vocabulary is seen as a decline of recognition when out of the standard or similar words. In this case, reconstructing the system in order to add or extend a range of vocabulary is a way to solve the problem. This paper propose configured Bhattacharyya algorithm standing by speech recognition learning model using the Bayesian methods which reflect parameter estimation upon the model configuration scalability. It is recognized corrected standard model based on a characteristic of the phoneme using the Bayesian methods for parameter estimation of the phoneme's data and Bhattacharyya algorithm for a similar model. By Bhattacharyya algorithm to configure recognition model evaluates a recognition performance. The result of applying the proposed method is showed a recognition rate of 97.3% and a learning curve of 1.2 seconds.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.5
• /
• pp.447-453
• /
• 2020

In this paper, we propose a system to extract effective speaker representations from a speech signal using a deep learning method. Based on the fact that speech signal contains identity unrelated information such as text content, emotion, background noise, and so on, we perform a training such that the extracted features only represent speaker-related information but do not represent speaker-unrelated information. Specifically, we propose an auto-encoder based disentanglement method that outputs both speaker-related and speaker-unrelated embeddings using effective loss functions. To further improve the reconstruction performance in the decoding process, we also introduce a discriminator popularly used in Generative Adversarial Network (GAN) structure. Since improving the decoding capability is helpful for preserving speaker information and disentanglement, it results in the improvement of speaker verification performance. Experimental results demonstrate the effectiveness of our proposed method by improving Equal Error Rate (EER) on benchmark dataset, Voxceleb1.

• The Journal of the Acoustical Society of Korea
• /
• v.22 no.5
• /
• pp.403-407
• /
• 2003

Recently, eigenvoice approach has been widely used for rapid speaker adaptation. However, even in the eigenvoice approach, Performance improvement using very small amount of adaptation data is relatively small in comparison with that using somewhat large adaptation data because the reliable estimation of weights of eigenvoice is difficult. In this paper, we propose a rapid speaker adaptation method based on eigenvoice using the weight distribution characteristics to improve the performance on a small adaptation data. In the Experimental results on vocabulary-independent word recognition task (using PBW 452 database), the weight threshold method alleviates the problem of relatively low performance for a tiny small adaptation data. When single adaptation word is used, word error rate is reduced about 9-18% by the weight threshold method.

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.2
• /
• pp.27-34
• /
• 2001

Utterance verification is used in variable vocabulary word recognition to reject the word that does not belong to in-vocabulary word or does not belong to correctly recognized word. Utterance verification is an important technology to design a user-friendly speech recognition system. We propose a new utterance verification algorithm for no-training utterance verification system based on the minimum verification error. First, using PBW (Phonetically Balanced Words) DB (445 words), we create no-training anti-phoneme models which include many PLUs(Phoneme Like Units), so anti-phoneme models have the minimum verification error. Then, for OOV (Out-Of-Vocabulary) rejection, the phoneme-based confidence measure which uses the likelihood between phoneme model (null hypothesis) and anti-phoneme model (alternative hypothesis) is normalized by null hypothesis, so the phoneme-based confidence measure tends to be more robust to OOV rejection. And, the word-based confidence measure which uses the phoneme-based confidence measure has been shown to provide improved detection of near-misses in speech recognition as well as better discrimination between in-vocabularys and OOVs. Using our proposed anti-model and confidence measure, we achieve significant performance improvement; CA (Correctly Accept for In-Vocabulary) is about 89％, and CR (Correctly Reject for OOV) is about 90％, improving about 15-21％ in ERR (Error Reduction Rate).

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.21 no.1
• /
• pp.144-150
• /
• 2017

Recently, deep learning is used for intelligent processing and accuracy improvement of data. It is formed calculation model composed of multi data processing layer that train the data representation through an abstraction of the various levels. A category of deep learning, convolution neural network is utilized in various research fields, which are human pose estimation, face recognition, image classification, speech recognition. When using the deep layer and lots of class, CNN that show a good performance on image classification obtain higher classification rate but occur the overfitting problem, when using a few data. So, we design the training network based on convolution neural network and trained our image data set for object classification in few class problem. The experiment show the higher classification rate of 7.06% in average than the previous networks designed to classify the object in 1000 class problem.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.30 no.6C
• /
• pp.505-511
• /
• 2005

Autonomic machines interacting with human should have capability to perceive the states of emotion and attitude through implicit messages for obtaining voluntary cooperation from their clients. Voice is the easiest and most natural way to exchange human messages. The automatic systems capable to understanding the states of emotion and attitude have utilized features based on pitch and energy of uttered sentences. Performance of the existing emotion recognition systems can be further improved withthe support of linguistic knowledge that specific tonal section in a sentence is related with the states of emotion and attitude. In this paper, we attempt to improve recognition rate of emotion by adopting such linguistic knowledge for Korean ending boundary tones into anautomatic system implemented using pitch-related features and multilayer perceptrons. From the results of an experiment over a Korean emotional speech database, the improvement of $4\%$ is confirmed.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2019.10a
• /
• pp.552-554
• /
• 2019

Deaf mutism (a hearing-impaired person and speech disorders) communicates using sign language. There are difficulties in communicating by voice. However, sign language can only be limited in communicating with people who know sign language because everyone doesn't use sign language when they communicate. In this paper, a finger language translation system is proposed and implemented as a means for the disabled and the non-disabled to communicate without difficulty. The proposed algorithm recognizes the finger language data by leap motion and self-learns the data using machine learning technology to increase recognition rate. We show performance improvement from the simulation results.

• The Journal of the Acoustical Society of Korea
• /
• v.14 no.2
• /
• pp.101-112
• /
• 1995

In order to construct a feature map-based phoneme classification system for speech recognition, two procedures are usually required. One is clustering and the other is labeling. In this paper, we present a phoneme classification system based on the Kohonen's Self-Organizing Feature Map (SOFM) for clusterer and labeler. It is known that the SOFM performs self-organizing process by which optimal local topographical mapping of the signal space and yields a reasonably high accuracy in recognition tasks. Consequently, SOFM can effectively be applied to the recognition of phonemes. Besides to improve the performance of the phoneme classification system, we propose the learning algorithm combined with the classical K-mans clustering algorithm in fine-tuning stage. In order to evaluate the performance of the proposed phoneme classification algorithm, we first use totaly 43 phonemes which construct six intra-class feature maps for six different phoneme classes. From the speaker-dependent phoneme classification tests using these six feature maps, we obtain recognition rate of $87.2\%$ and confirm that the proposed algorithm is an efficient method for improvement of recognition performance and convergence speed.