• Speech Sciences
• /
• v.8 no.4
• /
• pp.17-33
• /
• 2001

In this paper, we implement the large variable vocabulary speech recognition system, which is characterized by no additional pre-training process and no limitation of recognized word list. We have designed the system in order to achieve the high recognition rate using the decision tree based state tying algorithm and in order to reduce the processing time using the gaussian selection based variable flooring algorithm, the limitation algorithm of the number of nodes and ENNS algorithm. The gaussian selection based variable flooring algorithm shows that it can reduce the total processing time by more than half of the recognition time, but it brings about the reduction of recognition rate. In other words, there is a trade off between the recognition rate and the processing time. The limitation algorithm of the number of nodes shows the best performance when the number of gaussian mixtures is a three. Both of the off-line and on-line experiments show the same performance. In our experiments, there are some differences of the recognition rate and the average recognition time according to the distinction of genders, speakers, and the number of vocabulary.

• MALSORI
• /
• no.59
• /
• pp.53-68
• /
• 2006

We report the evaluation results of the Korean speech recognition platform called ECHOS. The platform has an object-oriented and reusable architecture so that researchers can easily evaluate their own algorithms. The platform has all intrinsic modules to build a large vocabulary speech recognizer: Noise reduction, end-point detection, feature extraction, hidden Markov model (HMM)-based acoustic modeling, cross-word modeling, n-gram language modeling, n-best search, word graph generation, and Korean-specific language processing. The platform supports both lexical search trees and finite-state networks. It performs word-dependent n-best search with bigram in the forward search stage, and rescores the lattice with trigram in the backward stage. In an 8000-word continuous speech recognition task, the platform with a lexical tree increases 40% of word errors but decreases 50% of recognition time compared to the HTK platform with flat lexicon. ECHOS reduces 40% of recognition errors through incorporation of cross-word modeling. With the number of Gaussian mixtures increasing to 16, it yields word accuracy comparable to the previous lexical tree-based platform, Julius.

• The Journal of the Acoustical Society of Korea
• /
• v.21 no.1E
• /
• pp.3-11
• /
• 2002

Broadcast news transcription is one of the hardest tasks in speech recognition because broadcast speech signals have much variability in speech quality, channel and background conditions. We developed a Korean broadcast news speech recognizer. We used a morpheme-based dictionary and a language model to reduce the out-of·vocabulary (OOV) rate. We concatenated the original morpheme pairs of short length or high frequency in order to reduce insertion and deletion errors due to short morphemes. We used a lexicon with multiple pronunciations to reflect inter-morpheme pronunciation variations without severe modification of the search tree. By using the merged morpheme as recognition units, we achieved the OOV rate of 1.7% comparable to European languages with 64k vocabulary. We implemented a hidden Markov model-based recognizer with vocal tract length normalization and online speaker adaptation by maximum likelihood linear regression. Experimental results showed that the recognizer yielded 21.8% morpheme error rate for anchor speech and 31.6% for mostly noisy reporter speech.

• ETRI Journal
• /
• v.32 no.5
• /
• pp.795-800
• /
• 2010

Embedding a large vocabulary speech recognition system in mobile devices requires a reduced acoustic model obtained by eliminating redundant model parameters. In conventional optimization methods based on the minimum description length (MDL) criterion, a binary Gaussian tree is built at each state of a hidden Markov model by iteratively finding and merging similar mixture components. An optimal subset of the tree nodes is then selected to generate a downsized acoustic model. To obtain a better binary Gaussian tree by improving the process of finding the most similar Gaussian components, this paper proposes a new distance measure that exploits the difference in likelihood values for cases before and after two components are combined. The mixture weight of Gaussian components is also introduced in the component merging step. Experimental results show that the proposed method outperforms MDL-based optimization using either a Kullback-Leibler (KL) divergence or weighted KL divergence measure. The proposed method could also reduce the acoustic model size by 50% with less than a 1.5% increase in error rate compared to a baseline system.

• Speech Sciences
• /
• v.10 no.1
• /
• pp.71-84
• /
• 2003

In the acoustic modeling for large vocabulary speech recognition, a sparse data problem caused by a huge number of context-dependent (CD) models usually leads the estimated models to being unreliable. In this paper, we develop a new clustering method based on the C45 decision-tree learning algorithm that effectively encapsulates the CD modeling. The proposed scheme essentially constructs a supervised decision rule and applies over the pre-clustered triphones using the C45 algorithm, which is known to effectively search through the attributes of the training instances and extract the attribute that best separates the given examples. In particular, the data driven method is used as a clustering algorithm while its result is used as the learning target of the C45 algorithm. This scheme has been shown to be effective particularly over the database of low unknown-context ratio in terms of recognition performance. For speaker-independent, task-independent continuous speech recognition task, the proposed method reduced the percent accuracy WER by 3.93% compared to the existing rule-based methods.

• Journal of Digital Convergence
• /
• v.13 no.1
• /
• pp.243-248
• /
• 2015

Recognition model is not defined when you configure a model, Been added to the model after model building awareness, Model a model of the clustering due to lack of recognition models are generated by modeling is causes the degradation of the recognition rate. In order to improve decision tree state tying modeling using parameter estimation of Bayesian method. The parameter estimation method is proposed Bayesian method to navigate through the model from the results of the decision tree based on the tying state according to the maximum probability method to determine the recognition model. According to our experiments on the simulation data generated by adding noise to clean speech, the proposed clustering method error rate reduction of 1.29% compared with baseline model, which is slightly better performance than the existing approach.

• MALSORI
• /
• no.46
• /
• pp.51-64
• /
• 2003

In order to produce high quality synthesized speech, it is very important to get an accurate grapheme-to-phoneme conversion and prosody model from texts using natural language processing. Robust preprocessing for non-Korean characters should also be required. In this paper, we analyzed Korean texts using a morphological analyzer, part-of-speech tagger and syntactic chunker. We present a new grapheme-to-phoneme conversion method for Korean using a hybrid method with a phonetic pattern dictionary and CCV (consonant vowel) LTS (letter to sound) rules, for unlimited vocabulary Korean TTS. We constructed a prosody model using a probabilistic method and decision tree-based method. The probabilistic method atone usually suffers from performance degradation due to inherent data sparseness problems. So we adopted tree-based error correction to overcome these training data limitations.

• Proceedings of the Acoustical Society of Korea Conference
• /
• autumn
• /
• pp.81-84
• /
• 1999

본 논문에서는 중규모급 단어인식기의 실시간 구현을 위한 무감독 단어집단화 알고리듬을 제안한다. 무감독 단어집단화는 인식대상 어휘 수가 많은 대용량 음성인식 시스템에서 대상 어휘 수를 줄여주는 역할을 하는 전처리기의 성격을 갖는다. 무감독 집단화를 위해 각 단어의 유$\cdot$무성음 고유의 특성을 잘 반영할 수 있는 특징 파라미터 5개를 사용하여 패턴 인식과 회귀분석에서 널리 사용되고 있는 분류$\cdot$회귀트리(Classification And Regression Tree)에 적용시키는 방법으로 접근하였고, 각 단어의 frame 수를 일정하게 n개로 분할(segment)하여 1개의 tree를 생성시키는 방법과 각 segment에 해당하는 tree를 생성시켜 segment들 사이의 교집합 성분으로 단어들을 집단화 하였다 실험결과 탐색 대상단어 22개에서 평균2.21개로 줄어 전체 대상 단어의 $10\%$만을 탐색하여 인식할 수 있는 방법을 제시할 수 있었다.

• The Journal of the Acoustical Society of Korea
• /
• v.23 no.2
• /
• pp.173-179
• /
• 2004

Due to the insufficiency of training data in large vocabulary continuous speech recognition, similar context dependent phones can be clustered by decision trees to share the data. When the decision trees are built and used to predict unseen triphones, a phonetic question set is required. The phonetic question set, which contains categories of the phones with similar co-articulation effects, is usually generated by phonetic or linguistic experts. This knowledge-based approach for generating phonetic question set, however, may reduce the homogeneity of the clusters. Moreover, the experts must adjust the question sets whenever the language or the PLU (phone-like unit) of a recognition system is changed. Therefore, we propose a data-driven method to automatically generate phonetic question set. Since the proposed method generates the phone categories using speech data distribution, it is not dependent on the language or the PLU, and may enhance the homogeneity of the clusters. In large vocabulary speech recognition experiments, the proposed algorithm has been found to reduce the error rate by 14.3%.

• The Journal of the Acoustical Society of Korea
• /
• v.25 no.7
• /
• pp.319-324
• /
• 2006

We Propose two new methods for non-keyword modeling to improve the performance of speaker- and vocabulary-independent keyword spotting system. The first method is decision tree clustering of monophone at the state level instead of monophone clustering method based on K-means algorithm. The second method is multi-state multiple mixture modeling at the syllable level rather than single state multiple mixture model for the non-keyword. To evaluate our method, we used the ETRI speech DB for training and keyword spotting test (closed test) . We also conduct an open test to spot 100 keywords with 400 sentences uttered by 4 speakers in an of fce environment. The experimental results showed that the decision tree-based state clustering method improve 28%/29% (closed/open test) than the monophone clustering method based K-means algorithm in keyword spotting. And multi-state non-keyword modeling at the syllable level improve 22%/2% (closed/open test) than single state model for the non-keyword. These results show that two proposed methods achieve the improvement of keyword spotting performance.