NaturalSpeech: End-to-End Text-to-Speech Synthesis With Human-Level Quality
Open accessThe DOI matches its Crossref record; 16 of the 16 references it cites were found in a registry or an index.
Abstract
Text-to-speech (TTS) has made rapid progress in both academia and industry in recent years. Some questions naturally arise that whether a TTS system can achieve human-level quality, how to define/judge that quality, and how to achieve it. In this paper, we answer these questions by first defining the human-level quality based on the statistical significance of subjective measure and introducing appropriate guidelines to judge it, and then developing a TTS system called NaturalSpeech that achieves human-level quality on benchmark datasets. Specifically, we leverage a variational auto-encoder (VAE) for end-to-end text-to-waveform generation, with several key modules to enhance the capacity of the prior from text and reduce the complexity of the posterior from speech, including phoneme pre-training, differentiable duration modeling, bidirectional prior/posterior modeling, and a memory mechanism in VAE. Experimental evaluations on the popular LJSpeech dataset show that our proposed NaturalSpeech achieves -0.01 CMOS (comparative mean opinion score) to human recordings at the sentence level, with Wilcoxon signed rank test at p-level p >> 0.05, which demonstrates no statistically significant difference from human recordings for the first time.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325 · cited 1×
“When using masked language modeling [46], we randomly mask some sup-phoneme tokens and their corresponding phoneme tokens and predict the masked phoneme and sup-phoneme at the same time.”
MethodAuto-Encoding Variational Bayes
Kingma, D. P., & Welling, M. (2013) · arXiv · doi:10.48550/arxiv.1312.6114 · Cited by 20,067
Decoupled Weight Decay Regularization
Loshchilov, I., & Hutter, F. (2017) · arXiv · doi:10.48550/arxiv.1711.05101 · Cited by 12,522
Neural Discrete Representation Learning
Oord, A. V. D., Vinyals, O., & Kavukcuoglu, K. (2017) · arXiv · doi:10.48550/arxiv.1711.00937 · Cited by 2,729
Taming Transformers for High-Resolution Image Synthesis
Esser, P., Rombach, R., & Ommer, B. (2020) · arXiv · doi:10.48550/arxiv.2012.09841 · Cited by 2,514
Taming Transformers for High-Resolution Image Synthesis
Esser, P., Rombach, R., & Ommer, B. (2021) · 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr46437.2021.01268 · Cited by 2,514
Zero-shot text-to-image generation
Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., & Sutskever, I. (2021) · arXiv · doi:10.48550/arxiv.2102.12092 · Cited by 2,417
Searching for activation functions
Ramachandran, P., Zoph, B., & Le, Q. V. (2017) · arXiv · doi:10.48550/arxiv.1710.05941 · Cited by 2,335 · cited 1×
“Conv1D ( · ) is one-dimensional convolution operation with layer normalization and Swish activation [47].”
Tacotron: towards end-to-end speech synthesis
Wang, Y., Skerry-Ryan, R. J., Stanton, D., Wu, Y., Weiss, R. J., Jaitly, N., Yang, Z., Xiao, Y., Chen, Z., Bengio, S., et al. (2017) · Interspeech 2017 · doi:10.21437/interspeech.2017-1452 · Cited by 1,746
NICE: Non-linear Independent Components Estimation
Dinh, L., Krueger, D., & Bengio, Y. (2014) · arXiv · doi:10.48550/arxiv.1410.8516 · Cited by 1,493 · cited 2×
“We choose a Flow model [25], [26], [27], [48] as the bidirectional prior/posterior module (denoted as θ bpp ) since it is easy to optimize and has a nice property of invertibility.”
Method“Waveglow: a flow-based generative network for speech synthesis,”
Prenger, R., Valle, R., & Catanzaro, B. (2019) · ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) · doi:10.1109/icassp.2019.8683143 · Cited by 1,004
Neural Machine Translation of Rare Words with Subword Units
Sennrich, R., Haddow, B., & Birch, A. (2015) · arXiv · doi:10.48550/arxiv.1508.07909 · Cited by 950
Neural Speech Synthesis with Transformer Network
Li, N., Liu, S., Liu, Y., Zhao, S., & Liu, M. (2019) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v33i01.33016706 · Cited by 754
Self-citationGenerating diverse high-fidelity images with vq-vae-2
Razavi, A., Oord, A. V. D., & Vinyals, O. (2019) · arXiv · doi:10.48550/arxiv.1906.00446 · Cited by 709 · cited 2×
“Inspired by image/video generation [21], [36], [37], [38], [39] that uses VQ-VAE [20], [40], [41] to compress high-dimensional image into low-dimensional representations to ease the generation [42], we leverage VAE [22] to compress high-dimensional speech x into frame-level representations z (i.e., z is sampled frompo…”
MethodA survey on neural speech synthesis
Tan, X., Qin, T., Soong, F., & Liu, T.-Y. (2021) · arXiv · doi:10.48550/arxiv.2106.15561 · Cited by 216
Self-citationVideoGPT: Video Generation using VQ-VAE and Transformers
Yan, W., Zhang, Y., Abbeel, P., & Srinivas, A. (2021) · arXiv · doi:10.48550/arxiv.2104.10157 · Cited by 156 · cited 1×
“Inspired by image/video generation [21], [36], [37], [38], [39] that uses VQ-VAE [20], [40], [41] to compress high-dimensional image into low-dimensional representations to ease the generation [42], we leverage VAE [22] to compress high-dimensional speech x into frame-level representations z (i.e.,…”
Method
Cited by
A Survey on the Detection and Impacts of Deepfakes in Visual, Audio, and Textual Formats
Mubarak, R., Alsboui, T., Alshaikh, O., Inuwa-Dutse, I., Khan, S., & Parkinson, S. (2023) · IEEE Access · doi:10.1109/access.2023.3344653 · Cited by 132
Domain Generalization via Aggregation and Separation for Audio Deepfake Detection
Xie, Y., Cheng, H., Wang, Y., & Ye, L. (2024) · IEEE Transactions on Information Forensics and Security · doi:10.1109/tifs.2023.3324724 · Cited by 59
Kim, J., & Fussell, S. R. (2025) · Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems · doi:10.1145/3706598.3714203 · Cited by 6
XMAD-Bench: Cross-Domain Multilingual Audio Deepfake Benchmark
Ciobanu, I.-P., Hiji, A.-I., Ristea, N.-C., Irofti, P., Rusu, C., & Ionescu, R. T. (2025) · arXiv · doi:10.48550/arxiv.2506.00462 · Cited by 1
Metadata
- Authors
- Tan, X., Chen, J., Liu, H., Cong, J., Zhang, C., Liu, Y., Wang, X., Leng, Y., Yi, Y., He, L., Zhao, S., Qin, T., Soong, F., & Liu, T.-Y.
- Venue
- IEEE Transactions on Pattern Analysis and Machine Intelligence
- Published
- 2024-06
- Volume
- 46
- Issue
- 6
- Pages
- 4234-4245
- ISSN
- 0162-8828 (ISSN-L)
- 2160-9292
- 1939-3539
- OpenAlex
- W4391020683
- Cited by
- 195
- References
- 16
- FWCI
- 25.89
- Type
- journal article
- Language
- en
- Open access
- green
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.