• Proceedings of the KSPS conference
• /
• 2007.05a
• /
• pp.327-329
• /
• 2007

This paper presents a FSN-based LVCSR system and it's application to the speech TV program guide. Unlike the most popular statistical language model-based system, we used FSN grammar based on the graph theory-based FSN optimization algorithm and knowledge-based advanced word boundary modeling. For the memory and latency efficiency, we implemented the dynamic pruning scheduling based on the histogram of active words and their likelihood distribution. We achieved a 10.7% word accuracy improvement with 57.3% speedup.

• Journal of Information Processing Systems
• /
• v.12 no.3
• /
• pp.410-421
• /
• 2016

In this paper, we propose a framework that attempts to incorporate landmarks into a segment-based Mandarin speech recognition system. In this method, landmarks provide boundary information and phonetic class information, and the information is used to direct the decoding process. To prove the validity of this method, two kinds of landmarks that can be reliably detected are used to direct the decoding process of a segment model (SM) based Mandarin LVCSR (large vocabulary continuous speech recognition) system. The results of our experiment show that about 30% decoding time can be saved without an obvious decrease in recognition accuracy. Thus, the potential of our method is demonstrated.

• Proceedings of the KSPS conference
• /
• 2006.11a
• /
• pp.24-27
• /
• 2006

This paper describes how a domain dependent pronunciation lexicon is generated and optimized for Korean large vocabulary continuous speech recognition(LVCSR). At the level of lexicon, pronunciation variations are usually modeled by adding pronunciation variants to the lexicon. We propose the criteria for selecting appropriate pronunciation variants in lexicon: (i) likelihood and (ii) frequency factors to select variants. Our experiment is conducted in three steps. First, the variants are generated with knowledge-based rules. Second, we generate a domain dependent lexicon which includes various numbers of pronunciation variants based on the proposed criteria. Finally, the WERs and RTFs are examined with each lexicon. In the experiment, 0.72% WER reduction is obtained by introducing the variants pruning criteria. Furthermore, RTF is not deteriorated although the average number of variants is higher than that of compared lexica.

• ETRI Journal
• /
• v.41 no.1
• /
• pp.109-116
• /
• 2019

This paper explores new techniques that are based on a hidden-layer linear transformation for fast speaker adaptation used in deep neural networks (DNNs). Conventional methods using affine transformations are ineffective because they require a relatively large number of parameters to perform. Meanwhile, methods that employ singular-value decomposition (SVD) are utilized because they are effective at reducing adaptive parameters. However, a matrix decomposition is computationally expensive when using online services. We propose the use of an extended diagonal linear transformation method to minimize adaptation parameters without SVD to increase the performance level for tasks that require smaller degrees of adaptation. In Korean large vocabulary continuous speech recognition (LVCSR) tasks, the proposed method shows significant improvements with error-reduction rates of 8.4% and 17.1% in five and 50 conversational sentence adaptations, respectively. Compared with the adaptation methods using SVD, there is an increased recognition performance with fewer parameters.

• The KIPS Transactions:PartB
• /
• v.10B no.4
• /
• pp.381-388
• /
• 2003

This paper presents a method of giving weights to garbage class clustering and Filler model to improve performance of keyword spotting system and a time-saving method of dialogue speech processing system for keyword spotting by calculating keyword transition probability through speech analysis of task domain users. The point of the method is grouping phonemes with phonetic similarities, which is effective in sensing similar phoneme groups rather than individual phonemes, and the paper aims to suggest five groups of phonemes obtained from the analysis of speech sentences in use in Korean morphology and in stock-trading speech processing system. Besides, task-subject Filler model weights are added to the phoneme groups, and keyword transition probability included in consecutive speech sentences is calculated and applied to the system in order to save time for system processing. To evaluate performance of the suggested system, corpus of 4,970 sentences was built to be used in task domains and a test was conducted with subjects of five people in their twenties and thirties. As a result, FOM with the weights on proposed five phoneme groups accounts for 85%, which has better performance than seven phoneme groups of Yapanel [1] with 88.5% and a little bit poorer performance than LVCSR with 89.8%. Even in calculation time, FOM reaches 0.70 seconds than 0.72 of seven phoneme groups. Lastly, it is also confirmed in a time-saving test that time is saved by 0.04 to 0.07 seconds when keyword transition probability is applied.

• Phonetics and Speech Sciences
• /
• v.10 no.4
• /
• pp.111-119
• /
• 2018

This paper proposes a method to improve speech recognition performance by extracting various pronunciations of the pseudo-morpheme unit from an eojeol unit corpus and generating a new recognition unit considering pronunciation variations. In the proposed method, we first align the pronunciation of the eojeol units and the pseudo-morpheme units, and then expand the pronunciation dictionary by extracting the new pronunciations of the pseudo-morpheme units at the pronunciation of the eojeol units. Then, we propose a new recognition unit that relies on pronunciation by tagging the obtained phoneme symbols according to the pseudo-morpheme units. The proposed units and their extended pronunciations are incorporated into the lexicon and language model of the speech recognizer. Experiments for performance evaluation are performed using the Korean speech recognizer with a trigram language model obtained by a 100 million pseudo-morpheme corpus and an acoustic model trained by a multi-genre broadcast speech data of 445 hours. The proposed method is shown to reduce the word error rate relatively by 13.8% in the news-genre evaluation data and by 4.5% in the total evaluation data.

• 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%.