{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/neural-speech-synthesis-with-transformer","title":"Neural Speech Synthesis with Transformer Network","arxiv_id":"1809.08895","date":"2018-09-19","proceeding":null,"authors":["Naihan Li","Shujie Liu","Yanqing Liu","Sheng Zhao","Ming Liu","Ming Zhou"],"abstract":"Although end-to-end neural text-to-speech (TTS) methods (such as Tacotron2)\nare proposed and achieve state-of-the-art performance, they still suffer from\ntwo problems: 1) low efficiency during training and inference; 2) hard to model\nlong dependency using current recurrent neural networks (RNNs). Inspired by the\nsuccess of Transformer network in neural machine translation (NMT), in this\npaper, we introduce and adapt the multi-head attention mechanism to replace the\nRNN structures and also the original attention mechanism in Tacotron2. With the\nhelp of multi-head self-attention, the hidden states in the encoder and decoder\nare constructed in parallel, which improves the training efficiency. Meanwhile,\nany two inputs at different times are connected directly by self-attention\nmechanism, which solves the long range dependency problem effectively. Using\nphoneme sequences as input, our Transformer TTS network generates mel\nspectrograms, followed by a WaveNet vocoder to output the final audio results.\nExperiments are conducted to test the efficiency and performance of our new\nnetwork. For the efficiency, our Transformer TTS network can speed up the\ntraining about 4.25 times faster compared with Tacotron2. For the performance,\nrigorous human tests show that our proposed model achieves state-of-the-art\nperformance (outperforms Tacotron2 with a gap of 0.048) and is very close to\nhuman quality (4.39 vs 4.44 in MOS).","url_abs":"http://arxiv.org/abs/1809.08895v3","url_pdf":"http://arxiv.org/pdf/1809.08895v3.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"neural-speech-synthesis-with-transformer","repo_url":"https://github.com/Munna-Manoj/Team6_FastSpeech2_TTS","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null},{"paper_slug":"neural-speech-synthesis-with-transformer","repo_url":"https://github.com/as-ideas/TransformerTTS","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":{"status":"ok","spdx":"NOASSERTION"}},{"paper_slug":"neural-speech-synthesis-with-transformer","repo_url":"https://github.com/choiHkk/Transformer-TTS","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"unanswered"}},{"paper_slug":"neural-speech-synthesis-with-transformer","repo_url":"https://github.com/soobinseo/transformer-tts","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null},{"paper_slug":"neural-speech-synthesis-with-transformer","repo_url":"https://github.com/tartunlp/transformertts","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":null},{"paper_slug":"neural-speech-synthesis-with-transformer","repo_url":"https://github.com/PaddlePaddle/PaddleSpeech","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":0,"framework":"paddle","reach":{"status":"ok","spdx":"Apache-2.0"}}],"tasks":[{"task_slug":"decoder","task_name":"Decoder"},{"task_slug":"machine-translation","task_name":"Machine Translation"},{"task_slug":"nmt","task_name":"NMT"},{"task_slug":"speech-synthesis","task_name":"Speech Synthesis"},{"task_slug":"text-to-speech","task_name":"Text to Speech"},{"task_slug":"text-to-speech-synthesis","task_name":"Text-To-Speech Synthesis"},{"task_slug":"text-to-speech-1","task_name":"text-to-speech"}],"methods":[{"method_slug":"absolute-position-encodings","method_name":"Absolute Position Encodings"},{"method_slug":"adam","method_name":"Adam"},{"method_slug":"attention","method_name":"Attention"},{"method_slug":"bpe","method_name":"BPE"},{"method_slug":"dense-connections","method_name":"Dense Connections"},{"method_slug":"dilated-causal-convolution","method_name":"Dilated Causal Convolution"},{"method_slug":"dropout","method_name":"Dropout"},{"method_slug":"label-smoothing","method_name":"Label Smoothing"},{"method_slug":"layer-normalization","method_name":"Layer Normalization"},{"method_slug":"linear-layer","method_name":"Linear Layer"},{"method_slug":"mixture-of-logistic-distributions","method_name":"Mixture of Logistic Distributions"},{"method_slug":"multi-head-attention","method_name":"Multi-Head Attention"},{"method_slug":"position-wise-feed-forward-layer","method_name":"Position-Wise Feed-Forward Layer"},{"method_slug":"relu","method_name":"ReLU"},{"method_slug":"residual-connection","method_name":"Residual Connection"},{"method_slug":"speed","method_name":"SPEED"},{"method_slug":"softmax","method_name":"Softmax"},{"method_slug":"transformer","method_name":"Transformer"},{"method_slug":"wavenet","method_name":"WaveNet"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/text-to-speech-synthesis-on-ljspeech","task":"Text-To-Speech Synthesis","dataset":"LJSpeech","model":"Transformer TTS (Mel + WaveGlow)","rank_in_archive_order":9,"of":16,"metrics":{"Audio Quality MOS":"3.88"},"uses_additional_data":true}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1809.08895","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}