• Journal of the Korea Society of Computer and Information
• /
• v.26 no.10
• /
• pp.1-7
• /
• 2021

Recently, since most of the research on correcting speech recognition errors is based on English, there is not enough research on Korean speech recognition. Compared to English speech recognition, however, Korean speech recognition has many errors due to the linguistic characteristics of Korean language, such as Korean Fortis and Korean Liaison, thus research on Korean speech recognition is needed. Furthermore, earlier works primarily focused on editorial distance algorithms and syllable restoration rules, making it difficult to correct the error types of Korean Fortis and Korean Liaison. In this paper, we propose a context-sensitive post-processing model of speech recognition using a LSTM-based sequence-to-sequence model and Bahdanau attention mechanism to correct Korean speech recognition errors caused by the pronunciation. Experiments showed that by using the model, the speech recognition performance was improved from 64% to 77% for Fortis, 74% to 90% for Liaison, and from 69% to 84% for average recognition than before. Based on the results, it seems possible to apply the proposed model to real-world applications based on speech recognition.

• MALSORI
• /
• no.61
• /
• pp.101-111
• /
• 2007

A modified speech recogniser have been proposed for automatic capitalisation generation to improve the readability of English speech recognition output. In this modified speech recogniser, every word in its vocabulary is duplicated: once in a de-caplitalised form and again in the capitalised forms. In addition its language model is re-trained on mixed case texts. In order to evaluate the performance of the proposed system, experiments of automatic capitalisation generation were performed for 3 hours of Broadcast News(BN) test data using the modified HTK BN transcription system. The proposed system produced an F-measure of 0.7317 for automatic capitalisation generation with an SER of 48.55, a precision of 0.7736 and a recall of 0.6942.

• Speech Sciences
• /
• v.10 no.1
• /
• pp.27-39
• /
• 2003

This study attempts to verify the reliability of automatically generated segment labels as compared to those obtained by conventional labelling by hand. First of all, an autolabeller is constructed using the standard HMM speech recognition technique. For evaluation, we compare the automatically generated labels with manually annotated labels for the same speech data. The comparison is performed by calculating the temporal difference between an autolabel boundary and its corresponding hand label boundary. When the mismatched duration between two labels falls within 10 msec, we consider the autolabel as correct. The results suggest that overall 78% of autolabels are correctly obtained. It is found that the boundary of obstruents is better aligned than that of sonorants and vowels. In case of stop sound classes, strong stops in manner-of-articulation wise and velar stops in place-of-articulation wise show better performance in boundary alignment. The result suggests that more phone-specific consideration is necessary to improve autosegmentation performance.

• Phonetics and Speech Sciences
• /
• v.15 no.3
• /
• pp.83-88
• /
• 2023

This study investigates the fine-tuning of large-scale Automatic Speech Recognition (ASR) models, specifically OpenAI's Whisper model, for domain-specific applications using the KsponSpeech dataset. The primary research questions address the effectiveness of targeted lexical item emphasis during fine-tuning, its impact on domain-specific performance, and whether the fine-tuned model can maintain generalization capabilities across different languages and environments. Experiments were conducted using two fine-tuning datasets: Set A, a small subset emphasizing specific lexical items, and Set B, consisting of the entire KsponSpeech dataset. Results showed that fine-tuning with targeted lexical items increased recognition accuracy and improved domain-specific performance, with generalization capabilities maintained when fine-tuned with a smaller dataset. For noisier environments, a trade-off between specificity and generalization capabilities was observed. This study highlights the potential of fine-tuning using minimal domain-specific data to achieve satisfactory results, emphasizing the importance of balancing specialization and generalization for ASR models. Future research could explore different fine-tuning strategies and novel technologies such as prompting to further enhance large-scale ASR models' domain-specific performance.

• Phonetics and Speech Sciences
• /
• v.5 no.2
• /
• pp.43-51
• /
• 2013

English native listeners have a tendency to treat strong syllables in a speech stream as the potential initial syllables of new words, since the majority of lexical words in English have a word-initial stress. The current study investigates whether Korean (L1) - English (L2) late bilinguals perceive strong syllables in English continuous speech as word onsets, as English native listeners do. In Experiment 1, word-spotting was slower when the word-initial syllable was strong, indicating that Korean listeners do not perceive strong syllables as word onsets. Experiment 2 was conducted in order to avoid any possibilities that the results of Experiment 1 may be due to the strong-initial targets themselves used in Experiment 1 being slower to recognize than the weak-initial targets. We employed the gating paradigm in Experiment 2, and measured the Isolation Point (IP, the point at which participants correctly identify a word without subsequently changing their minds) and the Recognition Point (RP, the point at which participants correctly identify the target with 85% or greater confidence) for the targets excised from the non-words in the two conditions of Experiment 1. Both the mean IPs and the mean RPs were significantly earlier for the strong-initial targets, which means that the results of Experiment 1 reflect the difficulty of segmentation when the initial syllable of words was strong. These results are consistent with Kim & Nam (2011), indicating that strong syllables are not perceived as word onsets for Korean listeners and interfere with lexical segmentation in English running speech.

• ETRI Journal
• /
• v.46 no.1
• /
• pp.48-58
• /
• 2024

This paper presents the development of language tutoring systems for nonnative speakers by leveraging advanced end-to-end automatic speech recognition (ASR) and proficiency evaluation. Given the frequent errors in non-native speech, high-performance spontaneous speech recognition must be applied. Our systems accurately evaluate pronunciation and speaking fluency and provide feedback on errors by relying on precise transcriptions. End-to-end ASR is implemented and enhanced by using diverse non-native speaker speech data for model training. For performance enhancement, we combine semisupervised and transfer learning techniques using labeled and unlabeled speech data. Automatic proficiency evaluation is performed by a model trained to maximize the statistical correlation between the fluency score manually determined by a human expert and a calculated fluency score. We developed an English tutoring system for Korean elementary students called EBS AI Peng-Talk and a Korean tutoring system for foreigners called KSI Korean AI Tutor. Both systems were deployed by South Korean government agencies.

• Journal of Korea Game Society
• /
• v.8 no.4
• /
• pp.107-114
• /
• 2008

The requests for speech-recognition for multi-language in field of game and the necessity of multi-language system, which expresses one phonetic model from many different kind of language phonetics, has been increased in field of game industry. Here upon, the research regarding development of multi-national language system which can express speeches, that is consist of various different languages, into only one lexical model is needed. In this paper is basic research for establishing integrated system from multi-language lexical model, and it shows the system which recognize Korean and English speeches into IPA(International Phonetic Alphabet). We focused on finding the IPA model which is satisfied with Korean and English phoneme one simutaneously. As a result, we could get the 90.62% of Korean speech-recognition rate, also 91.71% of English speech-recognition rate.

• Proceedings of the KSPS conference
• /
• 2003.05a
• /
• pp.168-171
• /
• 2003

This study is about modeling pronunciation dictionary necessary for PLU(phoneme like unit) level word recognition. The recognition of nonnative speakers' pronunciation enables an automatic diagnosis and an error detection which are the core of English pronunciation tutoring system. The above system needs two pronunciation dictionaries. One is for representing standard English pronunciation. The other is for representing Korean speakers' English Pronunciation. Both dictionaries are integrated to generate pronunciation networks for variants.

• Phonetics and Speech Sciences
• /
• v.2 no.1
• /
• pp.59-68
• /
• 2010

This study addresses the issue of whether Korean(L1)-English(L2) non-proficient bilinguals are affected by the native coda-neutralization process when recognizing words in English continuous speech. Korean phonological rules require that if liaison occurs between 'words', then coda-neutralization process must come before the liaison process, which results in liaison-consonants being coda-neutralized ones such as /b/, /d/, or /g/, rather than non-neutralized ones like /p/, /t/, /k/, /$t{\int}$/, /$d_{\Im}$/, or /s/. Consequently, if Korean listeners apply their native coda-neutralization rules to English speech input, word detection will be easier when coda-neutralized consonants precede target words than when non-neutralized ones do. Word-spotting and word-monitoring tasks were used in Experiment 1 and 2, respectively. In both experiments, listeners detected words faster and more accurately when vowel-initial target words were preceded by coda-neutralized consonants than when preceded by coda non-neutralized ones. The results show that Korean listeners exploit their native phonological process when processing English, irrespective of whether the native process is appropriate or not.

• Proceedings of the Acoustical Society of Korea Conference
• /
• spring
• /
• pp.99-102
• /
• 2000

In the case of making large vocabulary speech recognition system, it is better to use the segment than the syllable or the word as the recognition unit. In this paper, we study on the basis of making biphone for Korean speech recognition. For experiments, we use the speech toolkit of OGI in U.S.A. The result shows that the recognition rate of the case in which the diphthong is established as a single unit is superior to that of the case in which the diphthong Is established as two units, i.e. a glide plus a vowel. And also, the recognition rate of the case in which the biphone is used as the recognition unit is better than that of the case in which the mono-phoneme is used.