StyleTTS 2: Towards Human-Level Text-to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models

In this paper, we present StyleTTS 2, a text-to-speech (TTS) model that leverages style diffusion and adversarial training with large speech language models (SLMs) to achieve human-level TTS synthesis. StyleTTS 2 differs from its predecessor by modeling styles as a latent random variable through diffusion models to generate the most suitable style for the text without requiring reference speech, achieving efficient latent diffusion while benefiting from the diverse speech synthesis offered by diffusion models. Furthermore, we employ large pre-trained SLMs, such as WavLM, as discriminators with our novel differentiable duration modeling for end-to-end training, resulting in improved speech naturalness. StyleTTS 2 surpasses human recordings on the single-speaker LJSpeech dataset and matches it on the multispeaker VCTK dataset as judged by native English speakers. Moreover, when trained on the LibriTTS dataset, our model outperforms previous publicly available models for zero-shot speaker adaptation. This work achieves the first human-level TTS on both single and multispeaker datasets, showcasing the potential of style diffusion and adversarial training with large SLMs. The audio demos and source code are available at https://styletts2.github.io/.

Paper

References (65)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer mqLA6/10 · confidence 5/52023-07-02

Summary

This work presents a GAN-based TTS model, as an improvement on its predecessor StyleTTS. Notably, it includes a diffusion-based style encoder, which is one major difference compared to its predecessor. In addition, the proposed model directly produce waveform, in contrast to spectrogram in its predecessor.

Strengths

1) The proposed method is sound; 2) The audio samples sound great with clear improvements over multiple baselines 3) Abundant evaluations are conducted with persuasive results, including ablation studies

Weaknesses

1) The proposed model is quite complicated 2) A few technical questions, see below

Questions

1) "[StyleTTS 2] As the first model to achieve human-level performance on both single and multispeaker datasets" (line 47) -- Aren't PnG BERT and VITS achieved so as well, as acknowledged in Sec 2? If there is a difference on the scope of the claim, it's important to make it clear here. Also note that PnG BERT reported both MOS and CMOS (line 77). 2) Sec 3.2.1 claims "E2E training optimizes all TTS system components for inference without relying on any fixed components", but still "before jointly optimizing all components, we first pre-train the acoustic modules along with the pitch extractor and text aligner" -- is this still end-to-end training? 3) Sec 5.2: It's not clear how the evaluation is done. Are the styles analyzed the prosodic style, the acoustic style, or both? Are the emotion labels based on text, or audio (which can diverge)? what's the emotion of the reference audio? what about other factors as "style", beside emotion? Figure 2(c) -- "loose clusters" -- these are highly overlap to be considered as (even loose) "clusters" 4) Table 4: "real-time factor (RTF) in second" -- RTF is agnostic to the time unit 5) "SLM adversarial training" on "OOD texts" -- how does this work together with the regular training? There would be two types of training data, one is text-speech pairs, one is text-only -- how are they coordinated in an end-to-end manner? Can the model benefit from using model text-only data? 6) Evaluation on OOD text -- given the proposed model is trained on the OOD text (with SLM adversarial training), is it still a fair comparison to baseline models? importantly, are the eval text seen during SLM adversarial training? 7) Figure 1(b): where does $s_a$ come from during inference?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

Any plan on releasing the source code and the trained checkpoints, especially given the claim on "StyleTTS 2 sets a new benchmark for TTS synthesis"?

Reviewer 4B597/10 · confidence 4/52023-07-05

Summary

In this paper, the paper proposes StyleTTS 2 for human-level text-to-speech, which is an end-to-end system with two-stage training (pretraining some modules and finetuning the whole network). Two main modules proposed in paper is style diffusion module to model the prosodic style information and WavLM adversarial module to improve the quality of synthesized waveforms. Moreover, the paper proposes an interesting module on duration prediction, which is differentiable and stable.

Strengths

1. The paper proposes several interesting modules, including new differentiable duration modeling method, new adversarial module with WavLM and style diffusion to model the prosodic style information, which can benefit the future research in TTS. 2. The experiments verify the effectiveness and efficiency of proposed method by comparing MOS and RTF with different baseline methods. 3. The ablation study shows the effectiveness of proposed methods.

Weaknesses

I do not see obvious weakness.

Questions

1. For the multi-speaker setting or the zero-shot setting, how is the inference details. I think it will be different with Figure 1b since reference audio is used in inference? 2. Will the code be open-sourced to help the future research in TTS?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

N/A

Reviewer ogQb7/10 · confidence 4/52023-07-05

Summary

The paper presents Style TTS 2, a new TTS model with three main innovations 1) diffusion of a latent variable to capture style (everything in the speech signal that’s not enumerated by a phone sequence t and speaker variable c), 2) the use of a pretrained speech language model (SLM) as a discriminator and 3) a novel formulation of a differentiable alignment function. The authors also enable a new joint training curriculum. These three innovations result in quality that surpasses existing available TTS approaches including VITS, VALL-E, Natural Speech and YouTTS.

Strengths

The paper is well written, and, for the most part, easy to follow. Each of the novel components are clearly described, well motivated and contextualized with other comparable approaches. Also, the quality of the samples is quite good.

Weaknesses

Given that zero-shot speaker adaptation is extremely ripe for misuse as Societal Impact and Potential Harmful Consequences per https://neurips.cc/public/EthicsGuidelines, I would think the risk and mitigation section of this presentation should be more than 2 sentences. Section 3.2.1 using less resources is highlighted as a contribution of Style TTS 2 – does this include the “phoneme-level BERT pretrained on extensive corpora of Wikipedia articles”? Similar question concerning Section 3.2.3 and the use of WavLM pre-trained on 94k hours of data? It seems like this assessment is limited to the resources required for the model itself, but does not include components like the discriminator and text embedding. Some areas to tighten the presentation and contribution in the upfront material of the paper. Introduction: the first paragraph establishes the areas of improvement for TTS as diverse, expressive speech, robustness to OOD text, and requirements of massive datasets for performing zero-shot TTS. In the last paragraph of the introduction the contributions are oriented about quality, and novel use of SLMs. However, these are inconsistent with the areas for improvement enumerated by the authors. It might be clearer to be specific about which open questions for in the first paragraph are going to be addressed in some form by the proposed technique. The limitations of StyleTTS are again not the same as those established in the introduction. Namely they are 1) a two-stage training process, 2) limited expressiveness and 3) reliance on reference speech hindering realtime application. StyleTTS 2 addresses these three limitations, unifying the training process, and reducing reliance of reference speech while improving OOD performance and expressiveness. While this is compact and consistent as an improvement ot StyleTTS, it does not address the broad limitations of TTS established by the authors. Section 3.2 and 3.2.1 one of the advantages of Style TTS 2 is described as an E2E training process that jointly optimizes all components. This is presented in contrast to STyle TTS which required a two-stage training process (Section 3.1). However, later in Section 3.2.1 (line 142 and Figure 1a) Style TTS2 uses a two-stage training pre-training and joint training process. If single stage training, or E2E training is meant to be an advantage and contribution of Style TTS 2, the distinction between its two-stage training and the two-stage training used by the original Style TTS needs to be more clearly defined.

Questions

What does it mean for synthetic samples to “surpass” human recordings? Not in the context of how this happens as a score, but what does it say about the evaluation itself if it’s possible for an artificial sample to be more “natural” than a human derived sample which is “natural” with metaphysical certitude. I appreciate the argument about unnatural segmentation of the audio in LJSpeech, but if this is the source of the “unnaturalness” in the test data, it’s maybe not worth making too much of a claim about its value as a quantitative evaluation. On line 151, the augmented style vector is defined as s = [s_p,s_d], but s_d is not previously defined. Should this be s_a? Section 3.2.3 - the generator loss being independent of ground truth is true of any GAN training, no? Is there something specific to the use of an SLM discriminator that enables this better than prior work training TTS with GAN objectives? Section 5.1 “setting a new standard for this dataset”. What is the referent of “this dataset”? Is it the evaluation set of 80 text samples from LIbriTTS and 40 Librivox utterances spoken by the speaker of LJSpeech described in Section 4?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

Very briefly in two sentences at the end of the conclusion. (My rating below is contingent on the outcome of an ethics review to assess if the limitations and discussion of ethics concerns are sufficient to comply with https://neurips.cc/public/EthicsGuidelines)

Reviewer 5CmP7/10 · confidence 4/52023-07-06

Summary

The authors of the paper have proposed a high-quality TTS model StyleTTS 2. Unlike previous StyleTTS work, StyleTTS 2 has a DDPM-based style generator, speech language model discriminators, and a differentiable aligner. All of these design decisions help improve the overall performance of the model. As the authors report, StyleTTS 2 sets new standards for TTS synthesis.

Strengths

I appreciate the hard work of the authors on this paper. It was a well-written and enjoyable read. It is a great challenge to further develop TTS models with high quality, as the current SOTA models are very close to human speeches. StyleTTS 2 optimizes all the components in the joint training stage. Evaluation is thorough enough to support the arguments with in-depth ablation studies. I have carefully checked the audio samples, they are promising.

Weaknesses

The training procedures are complex. It requires several training stages, such as pre-training the acoustic modules, retraining the speech language models, and joint training. It is unclear whether the speech language models (SLMs) are trained on audio clips (rather than the whole-length audio). If so, how do you learn semantic information from short audio clips? It is also unclear whether the MPD and MRD still exist in the jointly training stage. I found MPD and MRD missing in Figure 1 (b). The overall novelty is limited, as many previous works integrate the diffusion modules and speech language model modules.

Questions

Do MPD and MRD still exist in the jointly training stage?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

No

Reviewer JmdN5/10 · confidence 4/52023-07-08

Summary

The paper introduces StyleTTS 2, a text-to-speech (TTS) model that leverages latent diffusion and adversarial training with SLMs. The most noteworthy thing is that the paper employs large pre-trained speech language models (SLMs), such as WavLM, as discriminators in the adversarial training process. This novel approach, combined with differentiable duration modeling for end-to-end training, improves the naturalness of the synthesized speech.

Strengths

1. It utilizes adversarial training with large speech language models (SLMs) in conjunction with differentiable duration modeling for end-to-end training. This approach improves the naturalness of synthesized speech. 2. Text encoder in StyleTTS 2 is separated into an Acoustic Text Encoder and a Prosodic Text Encoder. The Prosodic Text Encoder is then used as a conditional input for the latent diffusion process. This design allows the Diffusion model to model the diverse prosody of different styles, and it can also be seen through the ablation study that latent diffusion is essential in this paper.

Weaknesses

1. The article fails to provide a clear explanation of how the prosodic style encoder ensures the extraction of style from the Mel spectrogram rather than other elements. The specific mechanism behind this process is not adequately detailed in the paper. While the article briefly mentions certain style diffusion methods, they do not sufficiently explain how the prosodic style is accurately derived from the Mel spectrogram. Further exploration is required to devise more effective encoder architectures and training mechanisms to ensure the precise extraction of prosodic style information. 2. The writing style of the article lacks coherence, making it challenging to grasp the main ideas and problem-solving strategies. The organization and presentation of the paper could have been improved. ----------- The author's response addressed some of my issues, so I raised my score.

Questions

None

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

1 poor

Contribution

2 fair

Limitations

Yes

Reviewer 5CmP2023-08-17

Thanks for the comments

Thanks for your comments. I believe this paper merits acceptance as it is a high-quality paper, with a significant contribution that would be of interest to the community.

Reviewer ogQb2023-08-18

thank you

Thank you very much for answering my questions! I enjoyed reading the work and hope my comments helped to strengthen its presentation. I'll leave the assessment of the ethics comments to the ethics reviewers.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC