def __getitem__(self, index):
sample = self.data[index]
if not self.load_precomputed_mel:
features = AudioSegment.segment_from_file(
sample["audio_filepath"],
n_segments=self.n_segments
if self.n_segments is not None else -1,
trim=self.trim,
)
features = torch.tensor(features.samples)
audio, audio_length = features, torch.tensor(
features.shape[0]).long()
return audio, audio_length
else:
features = self.featurizer.process(sample["audio_filepath"],
trim=self.trim)
audio, audio_length = features, torch.tensor(
features.shape[0]).long()
mel = torch.load(sample["mel_filepath"])
frames = math.ceil(self.n_segments / self.hop_length)
if len(audio) > self.n_segments:
start = random.randint(0, mel.shape[1] - frames - 2)
mel = mel[:, start:start + frames]
audio = audio[start * self.hop_length:(start + frames) *
self.hop_length]
else:
mel = torch.nn.functional.pad(mel, (0, frames - mel.shape[1]))
audio = torch.nn.functional.pad(
audio, (0, self.n_segments - len(audio)))
return audio, len(audio), mel
def __getitem__(self, index):
"""
Given a index, returns audio and audio_length of the corresponding element. Audio clips of n_segments are
randomly chosen if the audio is longer than n_segments.
"""
example = self.collection[index]
features = AudioSegment.segment_from_file(example.audio_file, n_segments=self.n_segments, trim=self.trim,)
features = torch.tensor(features.samples)
audio, audio_length = features, torch.tensor(features.shape[0]).long()
truncate = audio_length % self.truncate_to
if truncate != 0:
audio_length -= truncate.long()
audio = audio[:audio_length]
return audio, audio_length