LibriMix

class torchaudio.datasets.LibriMix(root: Union[str, Path], subset: str = 'train-360', num_speakers: int = 2, sample_rate: int = 8000, task: str = 'sep_clean', mode: str = 'min')

LibriMix [Cosentino et al., 2020] dataset.

Parameters:

• root (str or Path) – The path to the directory where the directory Libri2Mix or Libri3Mix is stored.

• subset (str, optional) – The subset to use. Options: ["train-360", "train-100", "dev", and "test"] (Default: "train-360").

• num_speakers (int, optional) – The number of speakers, which determines the directories to traverse. The Dataset will traverse s1 to sN directories to collect N source audios. (Default: 2)

• sample_rate (int, optional) – Sample rate of audio files. The sample_rate determines which subdirectory the audio are fetched. If any of the audio has a different sample rate, raises ValueError. Options: [8000, 16000] (Default: 8000)

• task (str, optional) – The task of LibriMix. Options: ["enh_single", "enh_both", "sep_clean", "sep_noisy"] (Default: "sep_clean")

• mode (str, optional) – The mode when creating the mixture. If set to "min", the lengths of mixture and sources are the minimum length of all sources. If set to "max", the lengths of mixture and sources are zero padded to the maximum length of all sources. Options: ["min", "max"] (Default: "min")

Note

The LibriMix dataset needs to be manually generated. Please check https://github.com/JorisCos/LibriMix

__getitem__

LibriMix.__getitem__(key: int) → Tuple[int, Tensor, List[Tensor]]

Load the n-th sample from the dataset.

Parameters:

key (int) – The index of the sample to be loaded

Returns:

Tuple of the following items;

int:

Sample rate

Tensor:

Mixture waveform

List of Tensors:

List of source waveforms

get_metadata

LibriMix.get_metadata(key: int) → Tuple[int, str, List[str]]

Get metadata for the n-th sample from the dataset.

Parameters:

key (int) – The index of the sample to be loaded

Returns:

Tuple of the following items;

int:

Sample rate

str:

Path to mixed audio

List of str:

List of paths to source audios