VoxGenesis: Unsupervised Discovery of Latent Speaker Manifold for Speech Synthesis

We would like to thank the reviewer for the clear and constructive feedback. We have addressed the questions as follows:

Q1 I think the novelty is somewhat limited. To me, the proposed method is quite similar to a VAE-based speaker encoder but with the encoder changed to a pre-trained speaker embedding module. The mapping network is almost the same as in StyleGAN and is quite similar in principle to AdaSpeech. My main takeaway is that a pre-trained speaker embedding model as the encoder of a VAE works well for generating unseen speakers, and NFA is a promising speaker embedding model compared to other pre-trained speaker embedding model.

A1 We believe the reviewer missed the main contribution of our paper. If we have to put the takeaway in one sentence. It would be “constraining speaker embedding distribution into a known form allows speaker generation and speaker attribution manipulation”. Our contribution is not in introducing NFA per se or training speaker embedding in a VAE encoder-like fashion. Rather we identify a principled way to do speaker generation, i.e. by constraining speaker embedding distribution to a known form. For the mapping network, we mainly used it to transform the speaker embedding from an isotropic Gaussian distribution into a non-isotropic one to allow latent manipulation. StyleGAN does not use it for latent manipulation.

Q2 The baselines compared are quite old by deep learning standards. There are a few follow-up works to TacoSpawn. For zero-shot VC there are many more recent works (e.g. ControlVC). The authors state VITS is the "the state of the art" Multi-Speaker TTS but it is by no means true in 2023.

A2 We have expanded our experiments to include more baselines and datasets. Specifically, we have now incorporated FastSpeech2 and StyleTTS, both are recent popular papers, into our comparisons for multi-speaker TTS. Additionally, we have included ControlVC as an additional voice conversion baseline to provide a broader perspective on our model's performance.

Comparison of Multi-Speaker TTS Performance on LibriTTS-R, Original LibriTTS, and VCTK, Featuring Benchmarks Against VITS, StyleTTS, and FastSpeech

Results of Voice Conversion Experiments on LibriTTS-R, Original LibriTTS, and VCTK Datasets

We have checked the follow-up works to Tacospawn. The most popular one is "Creating new voices using normalizing flows" [1]. After reading the paper in detail, we found that their way to do speaker generation is entirely the same as Tacospawn. Therefore, we believe it is less interesting to include it. We are happy to include an additional speaker generation baselines if the reviewer has suggestions.

[1] Bilinski, P., Merritt, T., Ezzerg, A., Pokora, K., Cygert, S., Yanagisawa, K., Barra-Chicote, R., Korzekwa, D. (2022) Creating New Voices using Normalizing Flows. Proc. Interspeech 2022

Q3 & Q4: It would be better to compare the self-supervised speech representation model for the robustness of your methods by replacing HuBert with other SSL models. Using HuBERT representation may induce a high WER. Replacing it with ContentVec may improve the pronunciation.

A3 & A4: Thanks for the suggestion on SSL module. We have included results from other SSL-based content extractors, such as wav2vec-BERT and Content-Vec. The impact on WER is obvious. We will incorporate these findings in our revised paper.

Q5: The semantic conditioned Transformation is utilized in many works. NANSY++ utilizes a time-varying timbre embedding and this is almost the same with this part.

A5: We would like to clarify that we do not claim the semantic conditioned Transformation as our innovation. Instead, our contribution lies in constraining speaker embeddings p(z) to be easily sampled before transformation.

Q6: The authors cited LibriTTS and LibriTTS-R together. Which dataset do you use to train the models? In my personal experience, using LibriTTS-R decreases the sample diversity.

We used LibriTTS-R to train our model. Because it is a newly released dataset and based on LibriTTS, we cite both papers in case some readers are not familiar with LibriTTS-R. Our observations align with the reviewer: using the original LibriTTS leads to greater diversity in generated speakers compared to models trained with LibriTTS-R. We will include the results on the original LibriTTS in the appendix.

Q7: According to ICLR policy, when using human subjects such as MOS, you may include the evaluation details in your paper.

A7: The evaluation process of the Mean Opinion Score (MOS) is in Section 4.2 of our initial submission. This section comprehensively covers the protocols followed during the human subject evaluation.

Thanks for the clear and constructive feedback. We have addressed the questions as follows:

Q1: The authors should have conducted more comparisons to evaluate the model performance. and the result is a little incremental.

A1: We have expanded our experiments to include more baselines and datasets. Specifically, we have now incorporated FastSpeech2 and StyleTTS into our comparisons for multi-speaker TTS. Additionally, we have included ControlVC as an additional voice conversion baseline to provide a broader perspective on our model's performance.

Comparison of Multi-Speaker TTS Performance on LibriTTS-R, Original LibriTTS, and VCTK, Featuring Benchmarks Against VITS, StyleTTS, and FastSpeech

Results of Voice Conversion Experiments on LibriTTS-R, Original LibriTTS, and VCTK Datasets

Performance wise we believe the improvement is no incremental. Comparing with TacoSpawn, our NFA-VoxGenesis MOS improvement is 0.68, which by any criterion is significant improvement. The discrepancy in MOS is also demonstrated in sample quality in the demo. For voice conversion, the MOS margin with SpeechResynthesis is 0.44, which is also quit considerable improvement. The improvement is relatively small only in multi-speaker TTS.

Q2: They utilize a NFA which was proposed in ICML 2023. The contribution is weak.

A2: Our primary contribution is not the utilization of NFA. We demonstrate that regular speaker embedding networks, as well as unsupervised contrastive trained embedding networks, can be effectively used for novel speaker generation and speaker editing. The crux of our approach is constraining the speaker distribution p(z) into a form that allows for sampling. We introduce a principled method for speaker generation. Additionally, our work contributes to style control, particularly in independently controlling pitch, tone, and emotion through latent space editing. This contrasts with Style Tokens [1], which have attribute entanglement issues and differs from Style-TTS [2] which rely on external references for emotion control. Our pitch control also stands apart from pitch-specific techniques like ControlVC [3] and Fastspeech2 [4], which use pitch features and a dedicated pitch encoders.

Reference:

[1] Wang, Yuxuan, et al. "Style tokens: Unsupervised style modeling, control and transfer in end-to-end speech synthesis." International conference on machine learning. PMLR, 2018.


[2] Li, Yinghao Aaron, Cong Han, and Nima Mesgarani. "Styletts: A style-based generative model for natural and diverse text-to-speech synthesis." arXiv preprint arXiv:2205.15439 (2022).


[3] Chen, Meiying, and Zhiyao Duan. "ControlVC: Zero-Shot Voice Conversion with Time-Varying Controls on Pitch and Rhythm." arXiv preprint arXiv:2209.11866 (2022).


[4] Ren, Yi, et al. "FastSpeech 2: Fast and High-Quality End-to-End Text to Speech." International Conference on Learning Representations. 2021.

Thanks for the clear and constructive feedback. We have addressed the reviewer's questions as follows:

Q1.While the idea is promising, the experimental results seem to be limited. Most of the performances are from the ablation studies.

Q2. The paper can be more curated. While it is well written paper, it slightly lacks in structure.

A: We have made efforts to improve the structure of our paper by expanding experiments with more datasets and comparison with relevant recent works like StyleTTS [1], ControlVC [2], and FastSpeech2 [3].

Comparison of Multi-Speaker TTS Performance on LibriTTS-R, Original LibriTTS, and VCTK, Featuring Benchmarks Against VITS, StyleTTS, and FastSpeech

Results of Voice Conversion Experiments on LibriTTS-R, Original LibriTTS, and VCTK Datasets

Reference:

[1] Li, Yinghao Aaron, Cong Han, and Nima Mesgarani. "Styletts: A style-based generative model for natural and diverse text-to-speech synthesis." arXiv preprint arXiv:2205.15439 (2022).


[2] Chen, Meiying, and Zhiyao Duan. "ControlVC: Zero-Shot Voice Conversion with Time-Varying Controls on Pitch and Rhythm." arXiv preprint arXiv:2209.11866 (2022).


[3] Ren, Yi, et al. "FastSpeech 2: Fast and High-Quality End-to-End Text to Speech." International Conference on Learning Representations. 2021.

Q1 Sec 3.1: "GAN has been the de facto choice for vocoders." This is a false claim and ignores a large arrays of active and important works in the community. There are other popular vocoders choices like WaveNet, WaveRnn, diffusion-based approaches etc.

A1: We do not ignore the large arrays of active and important works in the community. Actually, we explicitly said so during the literature review section.

“Given the necessity for vocoders in both VC and TTS to invert the spectrogram, there has been a significant effort to improve them. This has led to the development of autoregressive, flow, GAN, and diffusion-based vocoders (Kong et al., 2020; Oord et al., 2016; Prenger et al., 2019; Chen et al., 2020). “


Q2 Sec 3.1: "A notable limitation in these models is ... learn to replicate the voices of the training or reference speakers rather than creating new voices." Another false claim. It improper to state for such an "limitation" because that is not the goal of the task of vocoders.

A2 We do not say the goal of the vocoder is to create the voice anywhere in our paper. We merely state a fact that the current GAN vocoder tries to replicate the voice of the training speaker instead of creating new voices.

Q3 Sec 3.1 "consequently, a conditional GAN is employed to transform a Gaussian distribution, rather than mapping speech features to waveforms". This is another improper comparison as what GAN does is to transfer the conditioning features (e.g. speech features) , rather than plain Gaussian noise, to waveforms.

A3 The whole novelty of our paper is to transform a known distribution (a Gaussian distribution) into a waveform, which are speaker embeddings. We have a hard time finding what is the improper comparison here.

Q4 Sec 3.1 "It is crucial, in the absence of Mel-spectrogram loss, for the discriminator to receive semantic tokens; otherwise, the generator could deceive the discriminator with intelligible speech." I don't see the connection.

Q4 Mel-spectrum loss is very important to GAN-based vocoder, without it the network does not receive conditional information, in the absence of it, we provided conditional information in the form of semantic tokens.

Q5 Sec 4.2 says speaker similarity is evaluated in a 3-point scale from 0 to 1, but Table 3 shows speaker similarity as 4.x and 3.x values.

A5 Table 3 shows speaker MOS instead of speaker similarity. We can not understand why the reviewer thinks it is speaker similarity. Table 3 explicitly writes speaker MOS. Both speaker similarity and Speaker MOS were explicitly defined in our paper.