• Speech Sciences
• /
• v.7 no.2
• /
• pp.193-208
• /
• 2000

This paper describes a multi-step method of intonation modeling for corpus-based Korean speech synthesizer. We selected 1833 sentences considering various syntactic structures and built a corresponding speech corpus uttered by a female announcer. We detected the pitch using laryngograph signals and manually marked the prosodic boundaries on recorded speech, and carried out the tagging of part-of-speech and syntactic analysis on the text. The detected pitch was separated into 3 frequency bands of low, mid, high frequency components which correspond to the baseline, the word tone, and the syllable tone. We predicted them using the CART method and the Viterbi search algorithm with a word-tone-dictionary. In the collected spoken sentences, 1500 sentences were trained and 333 sentences were tested. In the layer of word tone modeling, we compared two methods. One is to predict the word tone corresponding to the mid-frequency components directly and the other is to predict it by multiplying the ratio of the word tone to the baseline by the baseline. The former method resulted in a mean error of 12.37 Hz and the latter in one of 12.41 Hz, similar to each other. In the layer of syllable tone modeling, it resulted in a mean error rate less than 8.3% comparing with the mean pitch, 193.56 Hz of the announcer, so its performance was relatively good.

• The Journal of the Acoustical Society of Korea
• /
• v.13 no.5
• /
• pp.51-58
• /
• 1994

These times many kinds of man-machine Interfaces using speech signal, speech recognizers or speech synthesizers, are proposed and utilized in practice. Especially speech synthesis system is widely used in our life. But its assessment method is still in its first stage. In this paper we propose a method to generate multi-syllable nonsense wordset for the purpose of synthetic speech assessment and applies the wordset to one commercial text-to-speech system. Some results about the experiment is suggested and it is verified that the method to generate a nonsense wordset can be used to assess the intelligibility of the synthesizer in phoneme level or in phonemic environmental level.

• The Journal of the Acoustical Society of Korea
• /
• v.22 no.2
• /
• pp.135-144
• /
• 2003

A method for the automatic segmentation of speech signals is described. The method is dedicated to the construction of a large database for a Text-To-Speech (TTS) synthesis system. The main issue of the work involves the refinement of an initial estimation of phone boundaries which are provided by an alignment, based on a Hidden Market Model(HMM). Multi-layer perceptron (MLP) was used as a phone boundary detector. To increase the performance of segmentation, a technique which individually trains an MLP according to phonetic transition is proposed. The optimum partitioning of the entire phonetic transition space is constructed from the standpoint of minimizing the overall deviation from hand labelling positions. With single speaker stimuli, the experimental results showed that more than 95% of all phone boundaries have a boundary deviation from the reference position smaller than 20 ms, and the refinement of the boundaries reduces the root mean square error by about 25%.

• The Journal of the Acoustical Society of Korea
• /
• v.13 no.1
• /
• pp.71-83
• /
• 1994

This paper deals with the improvement of the synthesized speech quality and naturality in the Korean TTS(Text-to-Speech) system. We had extracted the parameters(table2) such as its amplitude, duration and pitch period in a syllable through the analysis of speech waveforms(table1) in the time domain and synthesized syllables using them. To the frequencies of the Korean pronunciation large vocabulary dictionary we had synthesized speeches selected 229 syllables such as V types are 19, CV types are 80. VC types are 30 and CVC types are 100. According to the 4 Korean syllable types from the data format dictionary(table3) we had tested each 15 syllables with the objective MOS(Mean Opinion Score) evaluation method about the 4 items i.e., intelligibility, clearness, loudness, and naturality after selecting random group without the knowledge of them. As the results of experiments the qualities of them are very clear and we can control the prosodic elements such as durations, accents and pitch periods (fig9, 10, 11, 12).

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.07a
• /
• pp.453-455
• /
• 2020

본 논문은 시각 장애인을 위한 감정 음성 자막 서비스를 생성하는 종단 간(end-to-end) 감정 음성 합성 시스템(emotional text-to-speech synthesis system, TTS)의 음성 합성 속도를 높이면서도 합성음의 음질을 향상시키는 방법을 제안한다. 기존에 사용했던 전역 스타일 토큰(Global Style Token, GST)을 이용한 감정 음성 합성 방법은 다양한 감정을 표현할 수 있는 장점을 갖고 있으나, 합성음을 생성하는데 필요한 시간이 길고 학습할 데이터의 동적 영역을 효과적으로 처리하지 않으면 합성음에 클리핑(clipping) 현상이 발생하는 등 음질이 저하되는 양상을 보였다. 이를 보안하기 위해 본 논문에서는 새로운 데이터 전처리 과정을 도입하였고 기존의 보코더(vocoder)인 웨이브넷(WaveNet)을 웨이브알엔엔(WaveRNN)으로 대체하여 생성 속도와 음질 측면에서 개선됨을 보였다.

• The Journal of Korea Robotics Society
• /
• v.13 no.1
• /
• pp.16-25
• /
• 2018

The talking head (TH) indicates an utterance face animation generated based on text and voice input. In this paper, we propose the generation method of TH with facial expression and intonation by speech input only. The problem of generating TH from speech can be regarded as a regression problem from the acoustic feature sequence to the facial code sequence which is a low dimensional vector representation that can efficiently encode and decode a face image. This regression was modeled by bidirectional RNN and trained by using SAVEE database of the front utterance face animation database as training data. The proposed method is able to generate TH with facial expression and intonation TH by using acoustic features such as MFCC, dynamic elements of MFCC, energy, and F0. According to the experiments, the configuration of the BLSTM layer of the first and second layers of bidirectional RNN was able to predict the face code best. For the evaluation, a questionnaire survey was conducted for 62 persons who watched TH animations, generated by the proposed method and the previous method. As a result, 77% of the respondents answered that the proposed method generated TH, which matches well with the speech.

• Journal of the Korea Computer Graphics Society
• /
• v.26 no.3
• /
• pp.49-59
• /
• 2020

In this paper, we propose a speech animation synthesis specialized in Korean through a rule-based co-articulation model. Speech animation has been widely used in the cultural industry, such as movies, animations, and games that require natural and realistic motion. Because the technique for audio driven speech animation has been mainly developed for English, however, the animation results for domestic content are often visually very unnatural. For example, dubbing of a voice actor is played with no mouth motion at all or with an unsynchronized looping of simple mouth shapes at best. Although there are language-independent speech animation models, which are not specialized in Korean, they are yet to ensure the quality to be utilized in a domestic content production. Therefore, we propose a natural speech animation synthesis method that reflects the linguistic characteristics of Korean driven by an input audio and text. Reflecting the features that vowels mostly determine the mouth shape in Korean, a coarticulation model separating lips and the tongue has been defined to solve the previous problem of lip distortion and occasional missing of some phoneme characteristics. Our model also reflects the differences in prosodic features for improved dynamics in speech animation. Through user studies, we verify that the proposed model can synthesize natural speech animation.

• The Journal of the Convergence on Culture Technology
• /
• v.5 no.4
• /
• pp.407-412
• /
• 2019

In this paper, we propose a technique for processing of speech recognition in korean words. Speech recognition is a technology that converts acoustic signals from sensors such as microphones into words or sentences. Most foreign languages have less difficulty in speech recognition. On the other hand, korean consists of vowels and bottom consonants, so it is inappropriate to use the letters obtained from the voice synthesis system. That improving the conventional structure speech recognition can the correct words recognition. In order to solve this problem, a new algorithm was added to the existing speech recognition structure to increase the speech recognition rate. Perform the preprocessing process of the word and then token the results. After combining the result processed in the Levenshtein distance algorithm and the hashing algorithm, the normalized words is output through the consonant comparison algorithm. The final result word is compared with the standardized table and output if it exists, registered in the table dose not exists. The experimental environment was developed by using a smartphone application. The proposed structure shows that the recognition rate is improved by 2% in standard language and 7% in dialect.

• The Journal of the Acoustical Society of Korea
• /
• v.25 no.5
• /
• pp.213-221
• /
• 2006

Phone segmentation of speech waveform is especially important for concatenative text to speech synthesis which uses segmented corpora for the construction of synthetic units. because the quality of synthesized speech depends critically on the accuracy of the segmentation. In the beginning. the phone segmentation was manually performed. but it brings the huge effort and the large time delay. HMM-based approaches adopted from automatic speech recognition are most widely used for automatic segmentation in speech synthesis, providing a consistent and accurate phone labeling scheme. Even the HMM-based approach has been successful, it may locate a phone boundary at a different position than expected. In this paper. we categorized adjacent phoneme pairs and analyzed the mismatches between hand-labeled transcriptions and HMM-based labels. Then we described the dominant error patterns that must be improved for the speech synthesis. For the experiment. hand labeled standard Korean speech DB from ETRI was used as a reference DB. Time difference larger than 20ms between hand-labeled phoneme boundary and auto-aligned boundary is treated as an automatic segmentation error. Our experimental results from female speaker revealed that plosive-vowel, affricate-vowel and vowel-liquid pairs showed high accuracies, 99%, 99.5% and 99% respectively. But stop-nasal, stop-liquid and nasal-liquid pairs showed very low accuracies, 45%, 50% and 55%. And these from male speaker revealed similar tendency.

• The Journal of the Convergence on Culture Technology
• /
• v.6 no.2
• /
• pp.509-514
• /
• 2020

The conventional TTS system consists of several modules, including text preprocessing, parsing analysis, grapheme-to-phoneme conversion, boundary analysis, prosody control, acoustic feature generation by acoustic model, and synthesized speech generation. But TTS system with deep learning is composed of Text2Mel process that generates spectrogram from text, and vocoder that synthesizes speech signals from spectrogram. In this paper, for the optimal Korean TTS system construction we apply Tacotron2 to Tex2Mel process, and as a vocoder we introduce the methods such as WaveNet, WaveRNN, and WaveGlow, and implement them to verify and compare their performance. Experimental results show that WaveNet has the highest MOS and the trained model is hundreds of megabytes in size, but the synthesis time is about 50 times the real time. WaveRNN shows MOS performance similar to that of WaveNet and the model size is several tens of megabytes, but this method also cannot be processed in real time. WaveGlow can handle real-time processing, but the model is several GB in size and MOS is the worst of the three vocoders. From the results of this study, the reference criteria for selecting the appropriate method according to the hardware environment in the field of applying the TTS system are presented in this paper.