chương 1 - Tách Related work ra khỏi chương 2 Dackground để tạo thành chương riêng. - Phân lích thêm về các nghiên cứu lên quan (chương 3) - Phân tích lý do xây dựng các thành phần của mô hình đề xuất (chương 4). - Chỉ tiết hơn phần thi nghiệm (chương 5). - Trình bảy lại một số nội đung lý thuyết được tham khảo, Ngày tháng năm 2022 Giảo viên hướng dẫn Tác giả luận văn CHTỦ TỊCH HỘI BỎNG 3 Experiments and Evaluations 5.1 Dataset kh rà be 5.2 Grapheme-to-phoneme converter 5.3 Results 6 Conclusion References Abstract Speech synthesis technology is an essential component of current human-computer interaction systems, which assists users in receiving the oulpul of intelligent ma- chincs more naturally aud intuitively, and has thus received incrcasing interest in recent years, The primary use of speech synthesis or text-to-speech technology is to translate a text into spoken speech automatically.
‘The current research fo- cus is the deep learning-based end-to-end speech synthesis technology. which has a more powerful modeling ability, This report proposcs a novel dccp leaming- based speech synthesis model called FastTacotron, which can resolve some issues of previous models, Experiments on the LSpeech dataset show that FastTacotron can converge in just a couple of hours of training using a single GPU. Mareover, the proposed model can accelerate the inference process and achieve high speech quality, Furthermore, our model also allows controlling the prosody of synthesized speech, thus can create expressive speech. Keywords: Deep learning, text-to-speech.
3 Experiments and Evaluations 5.1 Dataset kh rà be 5.2 Grapheme-to-phoneme converter 5.3 Results 6 Conclusion References Declaration of Authorship and Topic Sentences Personal Information © Fullname: LAM XUAN TI + Email: thu.yn * Class: Data Science + Tel: 098 994 7001 + Program: Full-time program * ‘This thesis is performed at: Department of Computer Science - School of Information and Communication Technology * This thesis is performed: from 22/03/2021 to 24/10/2021 Goals of the Thesis © Proposing @ novel neural netwerk for Speoch Synthesis, * Conducting experiments and evaluating the proposed model. Main Tasks of the Thesis * Introduce the Spocch Synthesis problem and review traditional approaches. for this problem. * Present machine learning and deep learning background for Speech Synthesis problem.
* Propose a novel neural network-based system for Speech Synthesis. * Implement experiments and evaluation. 3 Experiments and Evaluations 5.1 Dataset kh rà be 5.2 Grapheme-to-phoneme converter 5.3 Results 6 Conclusion References * Conclude and oulline future developments Declaration of Authorship 1 - Lam Xuan ‘hu - hereby warrant that the work and presentation in this thesis are performed by myself under the supervision of Dr. Dinh Vier Sang.
All results prescuted in this thesis arc truthful and acc not copied from any other works. Hanoi, 24th Nav 2021 Auihor Lam Xuan Tha Attestation of the Supervisor on the Fulfillment of the Requirements Hanoi. 24th Nov 2021 Supervisor D Dinh Viet Sang cà Table of Contents Declaration of Authorship and ‘Topic Sentences Acknowledgement Abstract Lists of Figures Lists of Tables 1 Introduction 2 Background 21 Machine learning. ` 23 ID Convolutional neural “nelworke 2.4 Recurrent neural nerworks.
Related work 31 Autoregressive models 31 Taeowon. Non-antoregressive models.3 PastSpeech 2 Proposed method 4.1 Pre-net and Post-net 42 LSTM. 443 Variation ‘predictor pee 43.4 Length regulator 44 Vocoder 6 ee ee ee List of Figures Speech synthesis or Text-to-speech: the artificial production of hu- "ÝỶ eee qa e ees Speech synthesis is used in a wide range of applications, such as assistive technology and multimedia. eee Machine learning: a new programming paradigm.
A neural nctwork is parameterized by its weights. bee A Joss function measures the quality of the nelwork’s output. The loss score is used as a feedback signal to adjust the weights. eee Recurrent neural network.
Fee The encoder-decoder model with additive attention mechanism [1]. An alignment graph.0 eee Transformer model architecture [2]. Scaled Dot-Product Allention (leM} and. Multi lead Auention (right) E TexIlo-spoech prow Tacotron model architecture [3].
‘Tacotron 2 model architecture [4] Eorwardlacotron model architecture [5] FastSpeech model architecture [6]. FastPitch model architecture [7]. FastSpecch 2 model architecture [8]. Model architecture of FastTacotron.
ee eee The CBHG module (-D convolution bank + highway network + bidirectional GRU) [3]. This CBHG is a power module for ox trating representations from sequence | 33 The ILSTM celL. 35 Duration/Pitch/Energy Predictor. The duration, pitch.
and energy pre- dictors all have a similar model structure (but different model pa- FAMCEELS) NA NT ens 36 Length Regulator [6]. This module is uscd to expand the length pf the phoneme sequence to match the length of mel-spoctrogram sequence, as well as to control the voice speed and part of prewaly MeIGAN model architecture [9]. > = MEI los and Pích lo. Duration loss and Energy loss.
nb TransformerI'TS model [10] Declaration of Authorship and Topic Sentences Personal Information © Fullname: LAM XUAN TI + Email: thu.yn * Class: Data Science + Tel: 098 994 7001 + Program: Full-time program * ‘This thesis is performed at: Department of Computer Science - School of Information and Communication Technology * This thesis is performed: from 22/03/2021 to 24/10/2021 Goals of the Thesis © Proposing @ novel neural netwerk for Speoch Synthesis, * Conducting experiments and evaluating the proposed model. Main Tasks of the Thesis * Introduce the Spocch Synthesis problem and review traditional approaches. for this problem. * Present machine learning and deep learning background for Speech Synthesis problem.
* Propose a novel neural network-based system for Speech Synthesis. * Implement experiments and evaluation. List of Figures Speech synthesis or Text-to-speech: the artificial production of hu- "ÝỶ eee qa e ees Speech synthesis is used in a wide range of applications, such as assistive technology and multimedia. eee Machine learning: a new programming paradigm.
A neural nctwork is parameterized by its weights. bee A Joss function measures the quality of the nelwork’s output. The loss score is used as a feedback signal to adjust the weights. eee Recurrent neural network.
Fee The encoder-decoder model with additive attention mechanism [1]. An alignment graph.0 eee Transformer model architecture [2]. Scaled Dot-Product Allention (leM} and. Multi lead Auention (right) E TexIlo-spoech prow Tacotron model architecture [3].
‘Tacotron 2 model architecture [4] Eorwardlacotron model architecture [5] FastSpeech model architecture [6]. FastPitch model architecture [7]. FastSpecch 2 model architecture [8]. Model architecture of FastTacotron.
ee eee The CBHG module (-D convolution bank + highway network + bidirectional GRU) [3]. This CBHG is a power module for ox trating representations from sequence | 33 The ILSTM celL. 35 Duration/Pitch/Energy Predictor. The duration, pitch.
and energy pre- dictors all have a similar model structure (but different model pa- FAMCEELS) NA NT ens 36 Length Regulator [6]. This module is uscd to expand the length pf the phoneme sequence to match the length of mel-spoctrogram sequence, as well as to control the voice speed and part of prewaly MeIGAN model architecture [9]. > = MEI los and Pích lo. Duration loss and Energy loss.
nb TransformerI'TS model [10] Abstract Speech synthesis technology is an essential component of current human-computer interaction systems, which assists users in receiving the oulpul of intelligent ma- chincs more naturally aud intuitively, and has thus received incrcasing interest in recent years, The primary use of speech synthesis or text-to-speech technology is to translate a text into spoken speech automatically. ‘The current research fo- cus is the deep learning-based end-to-end speech synthesis technology. which has a more powerful modeling ability, This report proposcs a novel dccp leaming- based speech synthesis model called FastTacotron, which can resolve some issues of previous models, Experiments on the LSpeech dataset show that FastTacotron can converge in just a couple of hours of training using a single GPU. Mareover, the proposed model can accelerate the inference process and achieve high speech quality, Furthermore, our model also allows controlling the prosody of synthesized speech, thus can create expressive speech.
Keywords: Deep learning, text-to-speech. Declaration of Authorship and Topic Sentences Personal Information © Fullname: LAM XUAN TI + Email: thu.yn * Class: Data Science + Tel: 098 994 7001 + Program: Full-time program * ‘This thesis is performed at: Department of Computer Science - School of Information and Communication Technology * This thesis is performed: from 22/03/2021 to 24/10/2021 Goals of the Thesis © Proposing @ novel neural netwerk for Speoch Synthesis, * Conducting experiments and evaluating the proposed model. Main Tasks of the Thesis * Introduce the Spocch Synthesis problem and review traditional approaches. for this problem.
* Present machine learning and deep learning background for Speech Synthesis problem. * Propose a novel neural network-based system for Speech Synthesis. * Implement experiments and evaluation. Table of Contents Declaration of Authorship and ‘Topic Sentences Acknowledgement Abstract Lists of Figures Lists of Tables 1 Introduction 2 Background 21 Machine learning.
` 23 ID Convolutional neural “nelworke 2.4 Recurrent neural nerworks. Related work 31 Autoregressive models 31 Taeowon. Non-antoregressive models.3 PastSpeech 2 Proposed method 4.1 Pre-net and Post-net 42 LSTM. 443 Variation ‘predictor pee 43.4 Length regulator 44 Vocoder 6 ee ee ee 3 Experiments and Evaluations 5.1 Dataset kh rà be 5.2 Grapheme-to-phoneme converter 5.3 Results 6 Conclusion References 3 Experiments and Evaluations 5.1 Dataset kh rà be 5.2 Grapheme-to-phoneme converter 5.3 Results 6 Conclusion References Acknowledgements 1 am extremely grateful to my supervisor, Dr.
Dinh Viet Sang, who gave me the golden opportunity to do this wonderful project on the lopic of Speech Synthesis, which also helped me in doing a lot of rescarch and I came to know about so many new things. It was a great privilege and honor to work and study under his guidance. 1 would also like to express my gratitude to my parents for their love, caring, and sacrifices for educaling and preparing me for my future, Finally, I would like to thank my fricnds for their immense support and help during this project. Without their help, completing this project would have been very difficult, Abstract Speech synthesis technology is an essential component of current human-computer interaction systems, which assists users in receiving the oulpul of intelligent ma- chincs more naturally aud intuitively, and has thus received incrcasing interest in recent years, The primary use of speech synthesis or text-to-speech technology is to translate a text into spoken speech automatically.
‘The current research fo- cus is the deep learning-based end-to-end speech synthesis technology. which has a more powerful modeling ability, This report proposcs a novel dccp leaming- based speech synthesis model called FastTacotron, which can resolve some issues of previous models, Experiments on the LSpeech dataset show that FastTacotron can converge in just a couple of hours of training using a single GPU. Mareover, the proposed model can accelerate the inference process and achieve high speech quality, Furthermore, our model also allows controlling the prosody of synthesized speech, thus can create expressive speech. Keywords: Deep learning, text-to-speech.
* Conclude and oulline future developments Declaration of Authorship 1 - Lam Xuan ‘hu - hereby warrant that the work and presentation in this thesis are performed by myself under the supervision of Dr. Dinh Vier Sang. All results prescuted in this thesis arc truthful and acc not copied from any other works. Hanoi, 24th Nav 2021 Auihor Lam Xuan Tha Attestation of the Supervisor on the Fulfillment of the Requirements Hanoi.
24th Nov 2021 Supervisor D Dinh Viet Sang cà * Conclude and oulline future developments Declaration of Authorship 1 - Lam Xuan ‘hu - hereby warrant that the work and presentation in this thesis are performed by myself under the supervision of Dr. Dinh Vier Sang. All results prescuted in this thesis arc truthful and acc not copied from any other works. Hanoi, 24th Nav 2021 Auihor Lam Xuan Tha Attestation of the Supervisor on the Fulfillment of the Requirements Hanoi.
24th Nov 2021 Supervisor D Dinh Viet Sang cà Acknowledgements 1 am extremely grateful to my supervisor, Dr.