Some questions

[Rongjiehuang]

Hi author, nice implementation! I wonder do we need some pre-trained models for self.emo_model = MinimalClassifier()?
Besides, I get audio samples with speaker identity inconsistent with reference audio.

[b04901014]

You won't need the pre-trained model for that explicitly. It should load the pretrained wav2vec2 by facebook from huggingface by default.

[Rongjiehuang]

class MinimalClassifier(pl.LightningModule):
    def __init__(self):
        super().__init__()
        self.wav2vec2 = Wav2vec2Wrapper(pretrain=False)

    def forward(self, x, length=None):
        reps = self.wav2vec2(x, length)
        return reps

As you see, the param pretrain is set to False in default, and I wonder whether it works but not typo?

[b04901014]

Sorry for the confusion. I borrowed the module Wav2vec2Wrapper from my another project. And that is that parameter is indicating if that module is used for pretraining or fine-tuning instead of whether it is pretrained or not. During pretraining it will output the masks for calculating the pretraining objective, which is not needed.

[Rongjiehuang]

Thanks for your clarification.