fix(meldataset): resampled audio may be shorter than (+1 more)

[andrewwhitecdw]

This PR addresses 2 issues in meldataset.py.

Fix 1

fix: resampled audio may be shorter than segment_size without padding

Fix: Replace:

                        if audio.shape[0] > self.segment_size:
                            # trim last elements to match self.segment_size (e.g., 16385 for 44khz downsampled to 24khz -> 16384)
                            audio = audio[: self.segment_size]

with:

                        if audio.shape[0] > self.segment_size:
                            # trim last elements to match self.segment_size (e.g., 16385 for 44khz downsampled to 24khz -> 16384)
                            audio = audio[: self.segment_size]
                        elif audio.shape[0] < self.segment_size:
                            audio = np.pad(
                                audio,
                                (0, self.segment_size - audio.shape[0]),
                                mode="constant",
                            )

Fix 2

fix: fine-tuning mel_start range off by one when frames equal segment

Fix: Replace:

                        mel_start = random.randint(0, mel.size(2) - frames_per_seg - 1)

with:

                        mel_start = random.randint(0, mel.size(2) - frames_per_seg)

Files changed

• meldataset.py