• The Journal of the Acoustical Society of Korea
• /
• v.15 no.2
• /
• pp.61-71
• /
• 1996

In this paper, dialogic triphone data base construction for triphone synthesis system is discussed. Particularly, in this work, dialogic speech data is collected from the broadcast media, and three different transcription steps are taken. Total 10 hours of speech data are collected. Among them, six hours of speech data are used for the triphone data base construction, and the rest four hours of data are reserved. Dialogic speech data base construction is far different from the reciting speech data base construction. This paper describes various steps that necessary for the dialogic triphone data base construction from collecting speech data to triphone unit labeling.

• Phonetics and Speech Sciences
• /
• v.7 no.1
• /
• pp.139-150
• /
• 2015

The data accumulated in this database will be used to develop a speaker identification system. This may also be applied towards, but not limited to, fields of phonetic studies, sociolinguistics, and language pathology. We plan to supplement the large-scale speech corpus next year, in terms of research methodology and content, to better answer the needs of diverse fields. The purpose of this study is to develop a speech corpus for standard Korean speech. For the samples to viably represent the state of spoken Korean, demographic factors were considered to modulate a balanced spread of age, gender, and dialects. Nine separate regional dialects were categorized, and five age groups were established from individuals in their 20s to 60s. A speech-sample collection protocol was developed for the purpose of this study where each speaker performs five tasks: two reading tasks, two semi-spontaneous speech tasks, and one spontaneous speech task. This particular configuration of sample data collection accommodates gathering of rich and well-balanced speech-samples across various speech types, and is expected to improve the utility of the speech corpus developed in this study. Samples from 639 individuals were collected using the protocol. Speech samples were collected also from other sources, for a combined total of samples from 1,012 individuals.

• Speech Sciences
• /
• v.10 no.2
• /
• pp.237-248
• /
• 2003

This paper compares the syllable structure of French and Korean analyzing the speech data of these two languages recorded during the actual speech. Reference to the syllable structure of French is made from F. Wioland's research data. As for the Korean data, the primary data are drawn from the 30-minute radio interview in which two male TV anchors in their early 60s talk to each other. The secondary source of the data is collected by having the primary data replicated by the two male announcers in their early 20's broadcasting in the university ra야o station of KAIST. With reference to the data collected in French and Korean, this paper provides the statistical frequency of each type of syllable structure in each language through the acoustic analysis of the spectrograms and renders a phonetic account of the characteristics of each syllable type in the two languages. Also discussed in this paper is the distributional condition in which each syllable structure is laid out in the speech context.

• Speech Sciences
• /
• v.10 no.1
• /
• pp.85-94
• /
• 2003

In this paper an algorithm to reduce the size of Text-to-Speech database is proposed. The algorithm is based on the characteristics of Korean phonemic units. From the initial database, a reduced phoneme unit set is induced by articulatory similarity of concatenating phonemes. Speech data is read by one female announcer for 1000 phonetically balanced sentences. All the recorded speech is then segmented by phoneticians. Total size of the original speech data is about 640 MB including laryngograph signal. To synthesize wave, RE-PSOLA (Residual-Excited Pitch Synchronous Overlap and Add Method) was used. The voice quality of synthesized speech was compared with original speech in terms of spectrographic informations and objective tests. The quality of the synthesized speech is not much degraded when the size of synthesis DB was reduced from 320 MB to 82 MB.

• Proceedings of the Korean Institute of Intelligent Systems Conference
• /
• 1993.06a
• /
• pp.1321-1325
• /
• 1993

This paper investigates the use of Fuzzy vector quantizer(FVQ) in speech synthesis. To compress speech data, we employ K-means algorithm to design codebook and then FVQ technique is used to analysize input speech vectors based on the codebook in an analysis part. In FVQ synthesis part, analysis data vectors generated in FVQ analysis is used to synthesize the speech. We have fined that synthesized speech quality depends on Fuzziness values in FVQ, and the optimum fuzziness values maximized synthesized speech SQNR are related with variance values of input speech vectors. This approach is tested on a sentence, and we compare synthesized speech by a convensional VQ with synthesized speech by a FVQ with optimum Fuzziness values.

• ETRI Journal
• /
• v.46 no.1
• /
• pp.71-81
• /
• 2024

Recent advances in deep learning for speech and visual recognition have accelerated the development of multimodal speech recognition, yielding many innovative results. We introduce a Korean audiovisual speech recognition corpus. This dataset comprises approximately 150 h of manually transcribed and annotated audiovisual data supplemented with additional 2000 h of untranscribed videos collected from YouTube under the Creative Commons License. The dataset is intended to be freely accessible for unrestricted research purposes. Along with the corpus, we propose an open-source framework for automatic speech recognition (ASR) and audiovisual speech recognition (AVSR). We validate the effectiveness of the corpus with evaluations using state-of-the-art ASR and AVSR techniques, capitalizing on both pretrained models and fine-tuning processes. After fine-tuning, ASR and AVSR achieve character error rates of 11.1% and 18.9%, respectively. This error difference highlights the need for improvement in AVSR techniques. We expect that our corpus will be an instrumental resource to support improvements in AVSR.

• Speech Sciences
• /
• v.9 no.1
• /
• pp.39-47
• /
• 2002

This paper proposes an effective method to remedy the acoustic modeling problem inherent in the usual log-normal Parallel Model Composition intended for achieving robust speech recognition. In particular, the Gaussian kernels under the prescribed log-normal PMC cannot sufficiently express the corrupted speech distributions. The proposed scheme corrects this deficiency by judiciously selecting the 'fairly' corrupted component and by re-estimating it as a mixture of two distributions using data-driven PMC. As a result, some components become merged while equal number of components split. The determination for splitting or merging is achieved by means of measuring the similarity of the corrupted speech model to those of the clean model and the noise model. The experimental results indicate that the suggested algorithm is effective in representing the corrupted speech distributions and attains consistent improvement over various SNR and noise cases.

• Phonetics and Speech Sciences
• /
• v.6 no.1
• /
• pp.77-83
• /
• 2014

This paper proposes a speaker and environment clustering method in order to overcome the degradation of the speech recognition performance caused by various noise and speaker characteristics. In this paper, instead of using the distance between Gaussian mixture model (GMM) weight vectors as in the Google's approach, the distance between the adapted mean vectors based on the modified maximum a posteriori (MAP) adaptation is used as a distance measure for vector quantization (VQ) clustering. According to our experiments on the simulation data generated by adding noise to clean speech, the proposed clustering method yields error rate reduction of 10.6% compared with baseline speaker-independent (SI) model, which is slightly better performance than the Google's approach.

• Proceedings of the Acoustical Society of Korea Conference
• /
• 1994.06a
• /
• pp.949-956
• /
• 1994

In this paper, we introduce a new speech synthesis method for Japanese and Korean arbitrary sentences using the natural speech data-base. Also, application of this method to a CAI system is discussed. In our synthesis method, a basic sentence and basic accent-phrases are selected from the data-base against a target sentence. Factors for those selections are phrase dependency structure (separation degree), number of morae, type of accent and phonemic labels. The target pitch pattern and phonemic parameter series are generated using those selected basic units. As the pitch pattern is generated using patterns which are directly extracted form real speech, it is expected to be more natural than any other pattern which is estimated by any model. Until now, we have examined this method on Japanese sentence speech and affirmed that the synthetic sound preserves human-like features fairly well. Now we extend this method to Korean sentence speech synthesis. Further more, we are trying to apply this synthesis unit to a CAI system.

• Speech Sciences
• /
• v.13 no.3
• /
• pp.7-18
• /
• 2006

Diagnosis of pathological voice is one of the important issues in biomedical applications of speech technology. This study focuses on the discrimination of voice disorder using HMM (Hidden Markov Model) for automatic detection between normal voice and vocal fold disorder voice. This is a non-intrusive, non-expensive and fully automated method using only a speech sample of the subject. Speech data from normal people and patients were collected. Mel-frequency filter cepstral coefficients (MFCCs) were modeled by HMM classifier. Different states (3 states, 5 states and 7 states), 3 mixtures and left to right HMMs were formed. This method gives an accuracy of 93.8% for train data and 91.7% for test data in the discrimination of normal and vocal fold disorder voice for sustained /a/.