• Speech Sciences
• /
• v.8 no.2
• /
• pp.83-90
• /
• 2001

We proposed a probabilistic approach of using syllable bigrams to the word-spacing problem. Syllable bigrams are extracted and the frequencies are calculated for the large corpus of 12 million words. Based on the syllable bigrams, we performed three experiments: (1) automatic word-spacing, (2) detection and correction of word-spacing errors for spelling checker, and (3) automatic insertion of a space at the end of line in the character recognition system. Experimental results show that the accuracy ratios are 97.7 percent, 82.1 percent, and 90.5%, respectively.

• Phonetics and Speech Sciences
• /
• v.8 no.2
• /
• pp.31-39
• /
• 2016

In emergency dispatching at 119 Command & Dispatch Center, some inconsistencies between the 'standard emergency aid system' and 'dispatch protocol,' which are both mandatory to follow, cause inefficiency in the dispatcher's performance. If an emergency dispatch system uses automatic speech recognition (ASR) to process the dispatcher's protocol speech during the case registration, it instantly extracts and provides the required information specified in the 'standard emergency aid system,' making the rescue command more efficient. For this purpose, we have developed a Korean large vocabulary continuous speech recognition system for 400,000 words to be used for the emergency dispatch system. The 400,000 words include vocabulary from news, SNS, blogs and emergency rescue domains. Acoustic model is constructed by using 1,300 hours of telephone call (8 kHz) speech, whereas language model is constructed by using 13 GB text corpus. From the transcribed corpus of 6,600 real telephone calls, call logs with emergency rescue command class and identified major symptom are extracted in connection with the rescue activity log and National Emergency Department Information System (NEDIS). ASR is applied to emergency dispatcher's repetition utterances about the patient information. Based on the Levenshtein distance between the ASR result and the template information, the emergency patient information is extracted. Experimental results show that 9.15% Word Error Rate of the speech recognition performance and 95.8% of emergency response detection performance are obtained for the emergency dispatch system.

• Phonetics and Speech Sciences
• /
• v.5 no.1
• /
• pp.99-108
• /
• 2013

Previous laboratory studies have shown that prosodic structures are encoded in the modulations of phonetic patterns of speech including suprasegmental as well as segmental features. In particular, effects of prosodic context on duration and intensity of syllables and words have been widely reported. Drawing on prosodically annotated large-scale speech data from the Buckeye corpus of conversational speech of American English, the current study attempted to examine whether and how prosodic prominence and phrase boundary of everyday conversational speech, as determined by a large group of ordinary listeners, are related to the phonetic realization of duration and intensity. The results showed that the patterns of word durations and intensities are influenced by prosodic structure. Closer examinations revealed, however, that the effects of prosodic prominence are not the same as those of prosodic phrase boundary. With regard to intensity measures, the results revealed the systematic changes in the patterns of overall RMS intensity near prosodic phrase boundary but the prominence effects are restricted to the nucleus. In terms of duration measures, both prosodic prominence and phrase boundary are the most closely related to the lengthening of the nucleus. Yet, prosodic prominence is more closely related to the lengthening of the onset while phrase boundary lengthens the coda duration more. The findings from the current study suggest that the phonetic realizations of prosodic prominence are different from those of prosodic phrase boundary, and speakers signal different prosodic structures through deliberate modulations of the internal phonetic structure of words and listeners attend to such phonetic variations.

• MALSORI
• /
• no.56
• /
• pp.225-235
• /
• 2005

Speech unit concatenation with a large database is presently the most popular method for speech synthesis. In this approach, the mismatches at the unit boundaries are unavoidable and become one of the reasons for quality degradation. This paper proposes an algorithm to reduce undesired discontinuities between the subsequent units. Optimal matching points are calculated in two steps. Firstly, the fullback-Leibler distance measurement is utilized for the spectral matching, then the unit sliding and the overlap windowing are used for the waveform matching. The proposed algorithm is implemented for the corpus-based unit concatenating Korean text-to-speech system that has an automatically labeled database. Experimental results show that our algorithm is fairly better than the raw concatenation or the overlap smoothing method.

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.4E
• /
• pp.31-38
• /
• 2001

A large-scale synthesis database for a unit selection based synthesis method usually retains redundant synthesis unit instances, which are useless to the synthetic speech quality. In this paper, to eliminate those instances from the synthesis database, we proposed a new pruning method called weighted vector quantization (WVQ). The WVQ reflects relative importance of each synthesis unit instance when clustering the similar instances using vector quantization (VQ) technique. The proposed method was compared with two conventional pruning methods through the objective and subjective evaluations of the synthetic speech quality: one to simply limit maximum number of instance, and the other based on normal VQ-based clustering. The proposed method showed the best performance under 50% reduction rates. Over 50% of reduction rates, the synthetic speech quality is not seriously but perceptibly degraded. Using the proposed method, the synthesis database can be efficiently reduced without serious degradation of the synthetic speech quality.

• MALSORI
• /
• no.64
• /
• pp.137-153
• /
• 2007

The pronunciation lexicon of a continuous speech recognition system should contain enough pronunciation variations to be used for building a search space large enough to contain a correct path, whereas the size of the pronunciation lexicon needs to be constrained for effective decoding and lower perplexities. This paper describes a procedure for selecting pronunciation variations to be included in the lexicon based on the frequencies of the corresponding phonetic rules observed in the training corpus. Likelihood of a phonetic rule's application is estimated using the observation frequency of the rule and is used to control the construction of a pronunciation lexicon. Experiments with various pronunciation lexica show that the proposed method is helpful to improve the speech recognition performance.

• MALSORI
• /
• no.51
• /
• pp.99-115
• /
• 2004

In this parer, we propose LM adaptation for broadcast news recognition. We collect information of recent articles from the internet on real time, make a recent small size LM, and then interpolate recent LM with a existing LM composed of existing large broadcast news corpus. We performed interpolation experiments to get the best type of articles from recent corpus because collected recent corpus is composed of articles which are related with test set, and which are unrelated. When we made an adapted LM using recent LM with similar articles to test set through Tf-Idf method and existing LM, we got the best result that ERR of pseudo-morpheme based recognition performance has 17.2 % improvement and the number of OOV has reduction from 70 to 27.

• The Journal of the Acoustical Society of Korea
• /
• v.28 no.6
• /
• pp.572-577
• /
• 2009

Large corpus-based concatenating Text-to-Speech (TTS) systems can generate natural synthetic speech without additional signal processing. To prune the redundant speech segments in a large speech segment DB, we can utilize a decision-tree based triphone clustering algorithm widely used in speech recognition area. But, the conventional methods have problems in representing the acoustic transitional characteristics of the phones and in applying context questions with hierarchic priority. In this paper, we propose a new clustering algorithm to downsize the speech DB. Firstly, three 13th order MFCC vectors from first, medial, and final frame of a phone are combined into a 39 dimensional vector to represent the transitional characteristics of a phone. And then the hierarchically grouped three question sets are used to construct the triphone trees. For the performance test, we used DTW algorithm to calculate the acoustic similarity between the target triphone and the triphone from the tree search result. Experimental results show that the proposed method can reduce the size of speech DB by 23% and select better phones with higher acoustic similarity. Therefore the proposed method can be applied to make a small sized TTS.

• Proceedings of the IEEK Conference
• /
• summer
• /
• pp.103-109
• /
• 2004

In NLP (Natural Language Processing) tasks, the highest difficulty which computers had to face with, is the built-in ambiguity of Natural Languages. To disambiguate it, formerly, they based on human-devised rules. Building such a complete rule-set is time-consuming and labor-intensive task whilst it doesn't cover all the cases. Besides, when the scale of system increases, it is very difficult to control that rule-set. So, recently, many NLP tasks have changed from rule-based approaches into corpus-based approaches with large annotated corpora. Corpus-based NLP tasks for such popular languages as English, French, etc. have been well studied with satisfactory achievements. In contrast, corpus-based NLP tasks for Vietnamese are at a deadlock due to absence of annotated training data. Furthermore, hand-annotation of even reasonably well-determined features such as part-of-speech (POS) tags has proved to be labor intensive and costly. In this paper, we present our building an annotated English-Vietnamese parallel aligned corpus named EVC to train for Vietnamese-related NLP tasks such as Word Segmentation, POS-tagger, Word Order transfer, Word Sense Disambiguation, English-to-Vietnamese Machine Translation, etc.

• 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%.