• Speech Sciences
• /
• v.5 no.1
• /
• pp.7-21
• /
• 1999

This paper proposes an improved training procedure in speech recognition based on the continuous density of the Hidden Markov Model (CDHMM). Of the three parameters (initial state distribution probability, state transition probability, output probability density function (p.d.f.) of state) governing the CDHMM model, we focus on the third parameter and propose an efficient algorithm that determines the p.d.f. of each state. It is known that the resulting CDHMM model converges to a local maximum point of parameter estimation via the iterative Expectation Maximization procedure. Specifically, we propose two independent algorithms that can be embedded in the segmental K -means training procedure by replacing relevant key steps; the adaptation of the number of mixture Gaussian p.d.f. and the initialization using the CDHMM parameters previously estimated. The proposed adaptation algorithm searches for the optimal number of mixture Gaussian humps to ensure that the p.d.f. is consistently re-estimated, enabling the model to converge toward the global maximum point. By applying an appropriate threshold value, which measures the amount of collective changes of weighted variances, the optimized number of mixture Gaussian branch is determined. The initialization algorithm essentially exploits the CDHMM parameters previously estimated and uses them as the basis for the current initial segmentation subroutine. It captures the trend of previous training history whereas the uniform segmentation decimates it. The recognition performance of the proposed adaptation procedures along with the suggested initialization is verified to be always better than that of existing training procedure using fixed number of mixture Gaussian p.d.f.

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.2
• /
• pp.3-9
• /
• 2001

In this paper, a random number generator for simulation of speech-waveform coders was designed. A random number generator having a desired probability density function and a desired power spectral density is discussed and experimental results are presented. The technique is based on Sondhi algorithm which consists of a linear filter and a memoryless nonlinearity. Several methods of obtaining memoryless nonlinearities for some typical continuous distributions are discussed. Sondhi algorithm is analyzed in the time domain using the diagonal expansion of the bivariate Gaussian probability density function. It is shown that the Sondhi algorithm gives satisfactory results when the memoryless nonlinearity is given in an antisymmetric form as in uniform, Cauchy, binary and gamma distribution. It is shown that the Sondhi algorithm does not perform well when the corresponding memoryless nonlinearity cannot be obtained analytically as in Student-t and F distributions, and when the memoryless nonlinearity can not be expressed in an antisymmetric form as in chi-squared and lognormal distributions.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.23 no.11
• /
• pp.1397-1406
• /
• 2019

In everyday mobile phone calls between the disabled and non-disabled people due to neurological impairment, the communication accuracy is often hindered by combining the accuracy of pronunciation due to the neurological impairment and the pronunciation features of the disabled. In order to improve this problem, the limiting method is MEMS (micro electro mechanical systems), which includes an induction line that artificially corrects difficult vocalization according to the oral characteristics of the language impaired by improving the word of out of vocabulary. mechanical System) Microphone device improvement. S/W improvement is decision tree with invert function, and improved matrix-vector rnn method is proposed considering continuous word characteristics. Considering the characteristics of H/W and S/W, a similar dictionary was created, contributing to the improvement of speech intelligibility for smooth communication.

• Journal of the Korean Institute of Telematics and Electronics C
• /
• v.35C no.11
• /
• pp.63-75
• /
• 1998

Since many co-articulation problems are occurring in continuous spoken Korean language, several researches use words as a basic recognition unit. Though the word unit can solve this problem, it requires much memory and has difficulty fitting an input speech in a word list. In this paper, we propose an hidden Markov model(HMM) based recognition model that is an interconnection network of word HMMs for a syntax of sentences. To match suitably the input sentence into the continuous word list in the network, we use a level building search algorithm. This system represents the large sentence set with a relatively small memory and also has good extensibility. The experimental result of an airplane reservation system shows that it is proper method for a practical recognition system.

• The Journal of the Acoustical Society of Korea
• /
• v.14 no.2E
• /
• pp.25-32
• /
• 1995

In this paper, we propose to use a parallel-branch subunit model for improved word recognition. The model is obtained by splitting off each subunit branch based on mixture component in continuous hidden Markov model(continuous HMM). According to simulation results, the proposed model yields higher recognition rate than the single-branch subunit model or the parallel-branch subunit model proposed by Rabiner et al[1]. We show that a proper combination of the number of mixture components and the number of branches for each subunit results in increased recognition rate. To study the recognition performance of the proposed algorithms, the speech material used in this work was a vocabulary with 1036 Korean words.

• The Journal of the Acoustical Society of Korea
• /
• v.19 no.6
• /
• pp.62-70
• /
• 2000

In usual language models(LMs), the probability has been estimated by selecting highly frequent words from a large text side database. However, in case of adopting LMs in a specific task, it is unnecessary to using the general method; constructing it from a large size tent, considering the various kinds of cost. In this paper, we propose a construction method of LMs using a small size text database in order to be used in specific tasks. The proposed method is efficient in increasing the low frequent words by applying same sentences iteratively, for it will robust the occurrence probability of words as well. We carried out continuous speech recognition(CSR) experiments on 200 sentences uttered by 3 speakers using LMs by iterative teaming(IL) in a air flight reservation task. The results indicated that the performance of CSR, using an IL applied LMs, shows an 20.4% increased recognition accuracy compared to those without it. This system, using the IL method, also shows an average of 13.4% higher recognition accuracy than the previous one, which uses context-free grammar(CFG), implying the effectiveness of it.

• The Journal of the Acoustical Society of Korea
• /
• v.16 no.3
• /
• pp.64-72
• /
• 1997

In this paper, study and experiments are performed for finding recognition unit fit which can be used in large vocabulary recognition system. Specifically, a phoneme that is currently used as recognition unit and a syllable in which Korean is well characterized are selected. From comparisons of recognition experiments, the study is performed whether a syllable can be considered as recognition unit of Korean recognition system. For report of an objective result of the comparison experiment, we collected speech data of a male speaker and processed them by hand-segmentation for phoneme boundary and labeling to construct speech database. And for training and recognition based on HMM, we used HTK (HMM Tool Kit) 2.0 of commercial tool from Entropic Co. to experiment in same condition. We applied two HMM model topologies, 3 emitting state of 5 state and 6 emitting state of 8 state, in Continuous HMM on training of each recognition unit. We also used 3 sets of PBW (Phonetically Balanced Words) and 1 set of POW(Phonetically Optimized Words) for training and another 1 set of PBW for recognition, that is "Speaker Dependent Medium Vocabulary Size Recognition." Experiments result reports that recognition rate is 95.65% in phoneme unit, 94.41% in syllable unit and decoding time of recognition in syllable unit is faster by 25% than in phoneme.

• The Journal of the Korea Contents Association
• /
• v.6 no.10
• /
• pp.9-16
• /
• 2006

In this paper, the acoustic model output of CSR(Continuous Speech Recognition) was used to analyze news data News database used in this experiment was consisted of 2,093 articles. Due to the low efficiency of language model, conventional Korean CSR is not appropriate to the analysis of news data. This problem could be handled successfully by introducing post-processing work of recognition result of acoustic model. The acoustic model more robust than language model in Korean environment. The result of post-processing work was made into KIF(Keyword information file). When threshold of acoustic model's output level was 100, 86.9% of whole target morpheme was included in post-processing result. At the same condition, applying length information based normalization, 81.25% of whole target morpheme was recognized. The purpose of normalization was to compensate long-length morpheme. According to experiment result, 75.13% of whole target morpheme was recognized KIF(314MB) had been produced from original news data(5,040MB). The decrease rate of absolute information met was approximately 93.8%.

• 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.29 no.1
• /
• pp.34-40
• /
• 2010

As internet spread widely, various service which use the internet have been provided. One of the service is a internet phone. Its usage is increasing by the advantage of cost. But it has a falling off in quality of speech. because it use packet switching method while existing telephone use circuit switching method. Although vocoder use PLC (Packet Loss Concealment) algorithm, it has a weakness of continuous packet loss. In this paper, we propose methods to improve a lowering in quality of speech under continuous loss of packet by using PLC algorithm used in advanced G.729 and G.711. The proposed methods are LP (Linear Prediction) parameter interpolation, excitation signal reconstruction and excitation signal gain reconstruction. As a result, the proposed method shows superior performance about 11%.