NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:8485
Title:MelGAN: Generative Adversarial Networks for Conditional Waveform Synthesis

Reviewer 1

Originality: The first work that trains GAN for spectrogram-to-waveform conversion without distillation. Quality: This paper suffers from a few critical issues. See comments below. Clarity: The experiment setting ups can be described with more details. Sec 3.2 and 3.4 is missing important information such as the datasets used for conducting the experiments. Significance: Although the quality of the proposed model remains unclear because of the previously mentioned critical issues, it’s a significant work because it’s the first GAN-based model for spectrogram-to-waveform conversion which seems to be working at some degree. Critical issues: 1. It’s significantly over-claimed: 1) claiming state-of-the-art for spectrogram-to-waveform conversion (line 6) with MOS 3.09 is surprising; many previous works are at a much higher level (e.g. 3.96 on the same dataset in, 4.5 in; 2) claiming state-of-the-art for text-to-speech synthesis (line 87) without comparison to strong baselines; 3) claiming “autoregressive models can be readily replaced with MelGAN decoder” (line 89, line 228) without necessary experiments (e.g. for TTS, no comparison to WaveNet and WaveRNN, no quantitative evaluation for multispeaker TTS; missing strong baselines) ; 4) claiming likely to work for linguistic-feature-to-waveform generation (footnote 1) without evidence. 2. Some experiment results seem in conflict. The MOS scores in a TTS setting up (3.88 in Table 3) is way higher than the MOS scores in a much easier (groundtruth-)spectrogram-to-waveform conversion setting up (3.09 in Table 2). Other comments: 1. Table 2: it will be very helpful to include the MOS for groundtruth audio, which serves as an anchor for comparison MOS scores (which is a subjective evaluation score). 2. Table 3: needs stronger baselines, such as Tacotron + WaveNet and Text-to-Mel + WaveNet, especially because the paper is claiming state-of-the-art. Also needs MOS on groundtruth. 3. It’s not clear what datasets are used for experiments in Sec 3.2 and 3.4. It will be helpful to make it clear. 4. Sec. 3.2: are the vocoder models trained on groundtruth spectrogram or predicted spectrogram? For best results, they should be trained on predicted spectrogram (see 5. Since spectrogram-to-waveform conversion is a strongly conditioned generative process, it’s interesting to know how much benefit GAN brings in this work. A baseline to compare with is to train only a Generator model with MSE loss (or other simple loss), without using Discriminator. 6. The author seems misunderstanding what VQ-VAE refers to (Sec. 3.4). VQ-VAE is a framework for learning discrete latent representation, which doesn’t include a WaveNet decoder (line 248) or other decoder (it’s typically paired with a decoder). It doesn't have to produce “downsampled” encoding either (line 247). 7. Line 93 “10 times faster than the fastest available model to date” -- please add a reference to the “fastest available model to date” in order to be clear which model is compared to. 8. Two papers should be covered as related work in the paragraph on line 73: and ============== Update: Thanks for the authors' response. The critical issues were addressed. I updated my score accordingly. However, the authors' response didn't address 6, 7, 8 in my detailed review. Please address them in the camera ready version. A couple more comments regarding the response: 1. Discrepancy in MOS scores between Table 3 and Table 2: comparison before training is converged is generally not trustworthy. Please justify why it's a proper comparison in the camera ready version. 2. The complete failure on the training without Discriminator sounds surprising. It may suggest headroom for improving in the design of the (strongly-conditional) Generator.

Reviewer 2

1. This paper successfully applied GAN objective on neural vocoder for speech synthesis. The model is moderately novel. 2. The paper present the model in a clear way. Equation/figures are accurate. Overall writing is high quality 3. The experimental results are missing the MOS of existing vocoders. For example, the authors should at least provide the MOS of autoregressive wavenet. For copy synthesis, the MOS of the proposed vocoder is only 3.09. The current state-of-the-art is above 4.0. So the authors can not claim the proposed vocoder as state-of-the-art. 4. Copy synthesized speech from the proposed vocoder contains clear artifacts. I don't think this vocoder can replace wavenet or wavernn.

Reviewer 3

UPDATE: having read the other reviews and the authors' rebuttal, I have decided not to change the review score. ---------- This paper proposes an adversarial model for mel-spectrogram inversion. This is an essential step in many modern audio generation pipelines. Mel-spectrogram inversion is challenging because this representation typically contains little to no phase information, and this information is essential to produce realistic sounding audio. While this has traditionally been tackled using signal processing methods (e.g. the Griffin-Lim phase reconstruction algorithm), recently several approaches using generative models have been proposed (e.g. using autoregressive models or flow-based models). The proposed use of an adversarial model in this setting is interesting: it is well known that adversarial models tend to forgo modelling all variations in the data in favour of focusing on a few modes, in order to be able to produce realistic examples given their limited capacity. This is actually a desirable property in the setting of e.g. text-to-speech, where we are simply after realistic conditional speech generation, and we don't care about capturing every possible variation in the speech signal corresponding to a given mel-spectrogram. However, success with adversarial models in the audio domain has been limited so far, and most of the literature has relied on likelihood-based models, so I think this is a timely paper. While the provided recordings of reconstructed and generated speech contain some audible artifacts (short hiccups / "doublespeak" which is characteristic of spectrogram inversion), the results are nevertheless impressive. My intuition is that they would need to be slightly better to be on par with production-level models such as parallel WaveNet, but they are remarkably close. The results in table 3 are impressive as well. The fidelity of the music samples is not so great, however. Remarks: - line 19: note that there is no intrinsic requirement for audio to have a sample rate of at least 16 kHz -- this rate just happens to be used quite commonly in literature for speech signals, because it is high enough for speech generation at a reasonable perceptual quality. - line 36: a very recent paper by Vasquez and Lewis ("MelNet: A Generative Model for Audio in the Frequency Domain") addresses the issue of robotic artifacts by using very high-resolution mel-spectrograms. Although this work wasn't publicly available at the time of submission, and is largely orthogonal in terms of its goals, I think it warrants a mention in the camera ready version of this work. - line 49: while it is true that WaveNet and other autoregressive models have trouble modelling information at time-scales larger than a few seconds, this does not seem particularly relevant in the context of mel-spectrogram inversion, and the criticism could arguably apply to almost any other model discussed in the paper. - Related to the previous comment: in some places, the paper would benefit from a bit more clarity with regards to which task is being considered: mel-spectrogram inversion, or audio generation in a general sense. Some statements only make sense in one context or the other, but it isn't always clear which context is meant. - lines 146-154: the motivation for the multiscale architecture refers to audio having structure at different levels. These "levels" are typically understood to be more than a factor of 2 or 4 apart though, so this motivation feels a bit out of place here. I think a link to wavelet-based processing of audio could perhaps be more appropriate here. - For the comparison in Table 1, it isn't clear at all whether the same hardware was used -- could you clarify? If not, these numbers would be considerably less meaningful, so this needs to be stated clearly. - Table 2 is excellent and clearly demonstrates the impact of the most important model design decisions.