• Journal of Korea Multimedia Society
• /
• v.6 no.3
• /
• pp.462-468
• /
• 2003

In a speech coding system using excitation source of voiced and unvoiced, it would be involved a distortion of speech quality in case coexist with a voiced and an unvoiced consonants in a frame. So, I propose TSIUVC(Transition Segment Including Unvoiced Consonant) searching and extraction method in order to uncoexistent with a voiced and unvoiced consonants in a frame. This paper present a new method of TSIUVC approximate-synthesis by using Least Mean Square and frequency band division. As a result, this method obtain a high quality approximation-synthesis waveforms within TSIUVC by using frequency information of 0.547KHz below and 2.813KHz above. The important thing is that the maximum error signal can be made with low distortion approximation-synthesis waveform within TSIUVC. This method has the capability of being applied to a new speech coding of Voiced/Silence/TSIUVC, speech analysis and speech synthesis.

• Phonetics and Speech Sciences
• /
• v.13 no.3
• /
• pp.65-70
• /
• 2021

This paper proposes a method for designing a large recording script for open domain English speech synthesis. For read-aloud style text, 12 domains and 294 sub-domains were designed using text contained in five different news media publications. For conversational style text, 4 domains and 36 sub-domains were designed using movie subtitles. The final script consists of 43,013 sentences, 27,085 read-aloud style sentences, and 15,928 conversational style sentences, consisting of 549,683 tokens and 38,356 types. The completed script is analyzed using four criteria: word coverage (type coverage and token coverage), high-frequency vocabulary coverage, phonetic coverage (diphone coverage and triphone coverage), and readability. The type coverage of our script reaches 36.86% despite its low token coverage of 2.97%. The high-frequency vocabulary coverage of the script is 73.82%, and the diphone coverage and triphone coverage of the whole script is 86.70% and 38.92%, respectively. The average readability of whole sentences is 9.03. The results of analysis show that the proposed method is effective in producing a large recording script for English speech synthesis, demonstrating good coverage in terms of unique words, high-frequency vocabulary, phonetic units, and readability.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.33 no.3C
• /
• pp.234-239
• /
• 2008

For the efficient coding and transmission, the speech signal can be classified into three distinctive classes: voiced, unvoiced, and transition classes. At low bit rate coding below 4 kbit/s, conventional sinusoidal transform coders synthesize speech of high quality for the purely voiced and unvoiced classes, whereas not for the transition class. The transition class including plosive sound and abrupt voiced-onset has the lack of periodicity, thus it is often classified and synthesized as the unvoiced class. In this paper, the efficient algorithm for the transition class detection is proposed, which demonstrates superior detection performance not only for clean speech but for noisy speech. For the detected transition frame, phase information is transmitted instead of magnitude information for speech synthesis. From the listening test, it was shown that the proposed algorithm produces better speech quality than the conventional one.

• Proceedings of the Acoustical Society of Korea Conference
• /
• 1998.06c
• /
• pp.159-162
• /
• 1998

Speech synthesis coding is classified into three categories: waveform coding, source coding and hybrid coding. To obtain the synthetic speech with high quality, the synthesis by waveform coding is desired. However, it is difficult to apply waveform coding to synthesis by syllable or phoneme unit, because it does not divide the speech into excitation and formant component. Thus it is required to alter the excitation in waveform coding for applying waveform coding to synthesis by rule. In this paper we propose a new pitch alteration method that minimizes the spectrum distortion by using the behavior of cepstrum. This method splits the spectrum of speech signal into excitation spectrum and formant spectrum and transforms the excitation spectrum into cepstrum domain. The pitch of excitation cepstrum is altered by zero insertion or zero deletion and the pitch altered spectrum is reconstructed in spectrum domain. As a result of performance test, the average spectrum distortion was below 2.29%.

• Phonetics and Speech Sciences
• /
• v.10 no.1
• /
• pp.39-48
• /
• 2018

A typical statistical parametric speech synthesis (text-to-speech, TTS) system consists of separate modules, such as a text analysis module, an acoustic modeling module, and a speech synthesis module. This causes two problems: 1) expert knowledge of each module is required, and 2) errors generated in each module accumulate passing through each module. An end-to-end TTS system could avoid such problems by synthesizing voice signals directly from an input string. In this study, we implemented an end-to-end Korean TTS system using Google's Tacotron, which is an end-to-end TTS system based on a sequence-to-sequence model with attention mechanism. We used 4392 utterances spoken by a Korean female speaker, an amount that corresponds to 37% of the dataset Google used for training Tacotron. Our system obtained mean opinion score (MOS) 2.98 and degradation mean opinion score (DMOS) 3.25. We will discuss the factors which affected training of the system. Experiments demonstrate that the post-processing network needs to be designed considering output language and input characters and that according to the amount of training data, the maximum value of n for n-grams modeled by the encoder should be small enough.

• Phonetics and Speech Sciences
• /
• v.4 no.3
• /
• pp.111-117
• /
• 2012

In this paper, we compare the performance of several speaker adaptation methods for a HMM-based Korean speech synthesis system with small amounts of adaptation data. According to objective and subjective evaluations, a hybrid method of constrained structural maximum a posteriori linear regression (CSMAPLR) and maximum a posteriori (MAP) adaptation shows better performance than other methods, when only five minutes of adaptation data are available for the target speaker. During the objective evaluation, we find that the duration models are insufficiently adapted to the target speaker as the spectral envelope and pitch models. To alleviate the problem, we propose the duration rectification method and the duration interpolation method. Both the objective and subjective evaluations reveal that the incorporation of the proposed two methods into the conventional speaker adaptation method is effective in improving the performance of the duration model adaptation.

• The Journal of the Acoustical Society of Korea
• /
• v.14 no.3
• /
• pp.49-60
• /
• 1995

This paper is a study on the design and implementation of the Korean Text-to-Speech system which is used for a small and simple system. In this paper, a parameter synthesis method is chosen for speech syntheiss method, we use PARCOR(PARtial autoCORrelation) coefficient which is one of the LPC analysis. And we use phoneme for synthesis unit which is the basic unit for speech synthesis. We use PARCOR, pitch, amplitude as synthesis parameter of voice, we use residual signal, PARCOR coefficients as synthesis parameter of unvoice. In this paper, we could obtain the 60% intelligibility by using the residual signal as excitation signal of unvoiced sound. The result of synthesis experiment, synthesis of a word unit is available. The controlling of phoneme duration is necessary for synthesizing of a sentence unit. For setting up the synthesis system, PC 486, a 70[Hz]-4.5[KHz] band pass filter for speech input/output, amplifier, and TMS320C30 DSP board was used.

• The Journal of the Convergence on Culture Technology
• /
• v.7 no.4
• /
• pp.675-681
• /
• 2021

With the recent growth of the AI speaker market, the demand for speech synthesis technology that enables natural conversation with users is increasing. Therefore, there is a need for a multi-speaker speech synthesis system that can generate voices of various tones. In order to synthesize natural speech, it is required to train with a large-capacity. high-quality speech DB. However, it is very difficult in terms of recording time and cost to collect a high-quality, large-capacity speech database uttered by many speakers. Therefore, it is necessary to train the speech synthesis system using the speech DB of a very large number of speakers with a small amount of training data for each speaker, and a technique for naturally expressing the tone and rhyme of multiple speakers is required. In this paper, we propose a technology for constructing a speaker encoder by applying the deep learning-based x-vector technique used in speaker recognition technology, and synthesizing a new speaker's tone with a small amount of data through the speaker encoder. In the multi-speaker speech synthesis system, the module for synthesizing mel-spectrogram from input text is composed of Tacotron2, and the vocoder generating synthesized speech consists of WaveNet with mixture of logistic distributions applied. The x-vector extracted from the trained speaker embedding neural networks is added to Tacotron2 as an input to express the desired speaker's tone.

• Journal of Korea Multimedia Society
• /
• v.7 no.9
• /
• pp.1233-1239
• /
• 2004

In a speech coding system using excitation source of voiced and unvoiced, it would be involved a distortion of speech qualify in case coexist with a voiced and an unvoiced consonants in a frame. This paper present a new multi-pulse coding method by using V/S/TSIUVC switching, individual pitch pulses and TSIUVC approximation-synthesis method in order to restrict a distortion of speech quality. The TSIUVC is extracted by using the zero crossing rate and individual pitch pulse. And the TSIUVC extraction rate was 91% for female voice and 96.2% for male voice respectively. The important thing is that the frequency information of 0.347kHz below and 2.813kHz above can be made with high quality synthesis waveform within TSIUVC. I evaluate the MPC use V/UV and the FBD-MPC use V/S/TSIUVC. As a result, I knew that synthesis speech of the FBD-MPC was better in speech quality than synthesis speech of the MPC.

• MALSORI
• /
• no.43
• /
• pp.137-150
• /
• 2002

Speech segmentation at phoneme level is important for corpus-based text-to-speech synthesis. In this paper, we examine acoustic modeling methods to improve the performance of automatic speech segmentation system based on Hidden Markov Model (HMM). We compare monophone and triphone models, and evaluate several model training approaches. In addition, we employ an energy-based postprocessing scheme to make correction of frequent boundary location errors between silence and speech sounds. Experimental results show that our system provides 71.3% and 84.2% correct boundary locations given tolerance of 10 ms and 20 ms, respectively.