• 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.13 no.4
• /
• pp.47-53
• /
• 2021

Autoregressive Text-to-Speech (TTS) models suffer from inference instability and slow inference speed. Inference instability occurs when a poorly predicted sample at time step t affects all the subsequent predictions. Slow inference speed arises from a model structure that forces the predicted samples from time steps 1 to t-1 to predict the sample at time step t. In this study, an end-to-end non-autoregressive fast text-to-speech model is suggested as a solution to these problems. The results of this study show that this model's Mean Opinion Score (MOS) is close to that of Tacotron 2 - WaveNet, while this model's inference speed and stability are higher than those of Tacotron 2 - WaveNet. Further, this study aims to offer insight into the improvement of non-autoregressive models.

• 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.

• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.14 no.2
• /
• pp.134-139
• /
• 2021

Currently, the autonomous vehicle market is commercializing a level 3 autonomous vehicle, but it still requires the attention of the driver. After the level 3 autonomous driving, the most notable aspect of level 4 autonomous vehicles is vehicle stability. This is because, unlike Level 3, autonomous vehicles after level 4 must perform autonomous driving, including the driver's carelessness. Therefore, in this paper, we propose the Emergency-notification and Driver-response Confirmation System(EDCS) for an autonomousvehicle safety that notifies the driver of an emergency situation and recognizes the driver's reaction in a situation where the driver is careless. The EDCS uses the emergency situation delivery module to make the emergency situation to text and transmits it to the driver by voice, and the driver response confirmation module recognizes the driver's reaction to the emergency situation and gives the driver permission Decide whether to pass. As a result of the experiment, the HMM of the emergency delivery module learned speech at 25% faster than RNN and 42.86% faster than LSTM. The Tacotron2 of the driver's response confirmation module converted text to speech about 20ms faster than deep voice and 50ms faster than deep mind. Therefore, the emergency notification and driver response confirmation system can efficiently learn the neural network model and check the driver's response in real time.

• Phonetics and Speech Sciences
• /
• v.13 no.3
• /
• pp.71-78
• /
• 2021

Most neural-network-based speech synthesis models utilize neural vocoders to convert mel-scaled spectrograms into high-quality, human-like voices. However, neural vocoders combined with mel-scaled spectrogram prediction models demand considerable computer memory and time during the training phase and are subject to slow inference speeds in an environment where GPU is not used. This problem does not arise in linear spectrogram prediction models, as they do not use neural vocoders, but these models suffer from low voice quality. As a solution, this paper proposes a Tacotron 2 and Transformer-based linear spectrogram prediction model that produces high-quality speech and does not use neural vocoders. Experiments suggest that this model can serve as the foundation of a high-quality text-to-speech model with fast inference speed.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2020.11a
• /
• pp.937-940
• /
• 2020

본 논문은 AI 기술을 기반으로 텍스트 스크립트를 자동으로 인식하고 영상 합성 기술을 응용하여 텍스트 정보를 시각화하는 AI 아나운서 소프트웨어 연구에 대하여 기술한다. 기존의 AI 기반 영상 정보 전달 서비스인 AI 앵커는 텍스트를 인식하여 영상을 합성하는데 오랜 시간이 필요하였으며, 특정 인물 이미지로만 영상 합성이 가능했기 때문에 그 용도가 제한적이었다. 본 연구에서 제안하는 방법은 Tacotron 으로 새로운 음성을 학습 및 합성하여, LRW 데이터셋으로 학습된 모델을 사용하여 자연스러운 영상 합성 체계를 구축한다. 단순한 얼굴 이미지의 합성을 개선하고 다채로운 이미지 제작을 위한 과정을 간략화하여 다양한 비대면 영상 정보 제공 환경을 구성할 수 있을 것으로 기대된다.

• 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.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2018.10a
• /
• pp.369-372
• /
• 2018

This paper suggests non-parallel-voice-conversion network conversing voice between unmapped voice pair as source voice and target voice. Conventional voice conversion researches used learning methods that minimize spectrogram's distance error. Not only these researches have some problem that is lost spectrogram resolution by methods averaging pixels. But also have used parallel data that is hard to collect. This research uses PPGs that is input voice's phonetic data and a GAN learning method to generate more clear voices. To evaluate the suggested method, we conduct MOS test with GMM based Model. We found that the performance is improved compared to the conventional methods.

• The Journal of the Acoustical Society of Korea
• /
• v.43 no.1
• /
• pp.9-18
• /
• 2024

The importance of active sonar systems is emerging due to the quietness of underwater targets and the increase in ambient noise due to the increase in maritime traffic. However, the low signal-to-noise ratio of the echo signal due to multipath propagation of the signal, various clutter, ambient noise and reverberation makes it difficult to identify underwater targets using active sonar. Attempts have been made to apply data-based methods such as machine learning or deep learning to improve the performance of underwater target recognition systems, but it is difficult to collect enough data for training due to the nature of sonar datasets. Methods based on mathematical modeling have been mainly used to compensate for insufficient active sonar data. However, methodologies based on mathematical modeling have limitations in accurately simulating complex underwater phenomena. Therefore, in this paper, we propose a sonar signal synthesis method based on a deep neural network. In order to apply the neural network model to the field of sonar signal synthesis, the proposed method appropriately corrects the attention-based encoder and decoder to the sonar signal, which is the main module of the Tacotron model mainly used in the field of speech synthesis. It is possible to synthesize a signal more similar to the actual signal by training the proposed model using the dataset collected by arranging a simulated target in an actual marine environment. In order to verify the performance of the proposed method, Perceptual evaluation of audio quality test was conducted and within score difference -2.3 was shown compared to actual signal in a total of four different environments. These results prove that the active sonar signal generated by the proposed method approximates the actual signal.