cckm/bigvgan_melspec

BigVGAN with different mel spectrogram input

These BigVGAN checkpoints are from continued training of https://huggingface.co/nvidia/bigvgan_v2_24khz_100band_256x, with the input mel spectrogram generated from this code from [vocos]:

class MelSpectrogramFeatures(FeatureExtractor):
    def __init__(self, sample_rate=24000, n_fft=1024, hop_length=256, n_mels=100, padding="center"):
        super().__init__()
        if padding not in ["center", "same"]:
            raise ValueError("Padding must be 'center' or 'same'.")
        self.padding = padding
        self.mel_spec = torchaudio.transforms.MelSpectrogram(
            sample_rate=sample_rate,
            n_fft=n_fft,
            hop_length=hop_length,
            n_mels=n_mels,
            center=padding == "center",
            power=1,
        )

    def forward(self, audio, **kwargs):
        if self.padding == "same":
            pad = self.mel_spec.win_length - self.mel_spec.hop_length
            audio = torch.nn.functional.pad(audio, (pad // 2, pad // 2), mode="reflect")
        mel = self.mel_spec(audio)
        features = safe_log(mel)
        return features

Training was done with segment_size=65536 (unchanged) and batch_size=24 (vs 32 from the Nvidia team). Final eval PESQ is 4.340 (vs 4.362 from the Nvidia checkpoint, on their own mel spectrogram code).