Shortcuts

torchaudio.datasets

All datasets are subclasses of torch.utils.data.Dataset i.e, they have __getitem__ and __len__ methods implemented. Hence, they can all be passed to a torch.utils.data.DataLoader which can load multiple samples parallelly using torch.multiprocessing workers. For example:

yesno_data = torchaudio.datasets.YESNO('.', download=True)
data_loader = torch.utils.data.DataLoader(yesno_data,
                                          batch_size=1,
                                          shuffle=True,
                                          num_workers=args.nThreads)

The following datasets are available:

All the datasets have almost similar API. They all have two common arguments: transform and target_transform to transform the input and target respectively.

CMUARCTIC

class torchaudio.datasets.CMUARCTIC(root: str, url: str = 'aew', folder_in_archive: str = 'ARCTIC', download: bool = False)[source]

Create a Dataset for CMU_ARCTIC.

Parameters
  • root (str) – Path to the directory where the dataset is found or downloaded.

  • url (str, optional) – The URL to download the dataset from or the type of the dataset to dowload. (default: "aew") Allowed type values are "aew", "ahw", "aup", "awb", "axb", "bdl", "clb", "eey", "fem", "gka", "jmk", "ksp", "ljm", "lnh", "rms", "rxr", "slp" or "slt".

  • folder_in_archive (str, optional) – The top-level directory of the dataset. (default: "ARCTIC")

  • download (bool, optional) – Whether to download the dataset if it is not found at root path. (default: False).

__getitem__(n: int) → Tuple[torch.Tensor, int, str, str][source]

Load the n-th sample from the dataset.

Parameters

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

Returns

(waveform, sample_rate, utterance, utterance_id)

Return type

tuple

COMMONVOICE

class torchaudio.datasets.COMMONVOICE(root: str, tsv: str = 'train.tsv', url: str = 'english', folder_in_archive: str = 'CommonVoice', version: str = 'cv-corpus-4-2019-12-10', download: bool = False)[source]

Create a Dataset for CommonVoice.

Parameters
  • root (str) – Path to the directory where the dataset is found or downloaded.

  • tsv (str, optional) – The name of the tsv file used to construct the metadata. (default: "train.tsv")

  • url (str, optional) – The URL to download the dataset from, or the language of the dataset to download. (default: "english"). Allowed language values are "tatar", "english", "german", "french", "welsh", "breton", "chuvash", "turkish", "kyrgyz", "irish", "kabyle", "catalan", "taiwanese", "slovenian", "italian", "dutch", "hakha chin", "esperanto", "estonian", "persian", "portuguese", "basque", "spanish", "chinese", "mongolian", "sakha", "dhivehi", "kinyarwanda", "swedish", "russian", "indonesian", "arabic", "tamil", "interlingua", "latvian", "japanese", "votic", "abkhaz", "cantonese" and "romansh sursilvan".

  • folder_in_archive (str, optional) – The top-level directory of the dataset.

  • version (str) – Version string. (default: "cv-corpus-4-2019-12-10") For the other allowed values, Please checkout https://commonvoice.mozilla.org/en/datasets.

  • download (bool, optional) – Whether to download the dataset if it is not found at root path. (default: False).

__getitem__(n: int) → Tuple[torch.Tensor, int, Dict[str, str]][source]

Load the n-th sample from the dataset.

Parameters

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

Returns

(waveform, sample_rate, dictionary), where dictionary is built from the TSV file with the following keys: client_id, path, sentence, up_votes, down_votes, age, gender and accent.

Return type

tuple

GTZAN

class torchaudio.datasets.GTZAN(root: str, url: str = 'http://opihi.cs.uvic.ca/sound/genres.tar.gz', folder_in_archive: str = 'genres', download: bool = False, subset: Optional[str] = None)[source]

Create a Dataset for GTZAN.

Note

Please see http://marsyas.info/downloads/datasets.html if you are planning to use this dataset to publish results.

Parameters
  • root (str) – Path to the directory where the dataset is found or downloaded.

  • url (str, optional) – The URL to download the dataset from. (default: "http://opihi.cs.uvic.ca/sound/genres.tar.gz")

  • folder_in_archive (str, optional) – The top-level directory of the dataset.

  • download (bool, optional) – Whether to download the dataset if it is not found at root path. (default: False).

  • subset (str, optional) – Which subset of the dataset to use. One of "training", "validation", "testing" or None. If None, the entire dataset is used. (default: None).

__getitem__(n: int) → Tuple[torch.Tensor, int, str][source]

Load the n-th sample from the dataset.

Parameters

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

Returns

(waveform, sample_rate, label)

Return type

tuple

LIBRISPEECH

class torchaudio.datasets.LIBRISPEECH(root: str, url: str = 'train-clean-100', folder_in_archive: str = 'LibriSpeech', download: bool = False)[source]

Create a Dataset for LibriSpeech.

Parameters
  • root (str) – Path to the directory where the dataset is found or downloaded.

  • url (str, optional) – The URL to download the dataset from, or the type of the dataset to dowload. Allowed type values are "dev-clean", "dev-other", "test-clean", "test-other", "train-clean-100", "train-clean-360" and "train-other-500". (default: "train-clean-100")

  • folder_in_archive (str, optional) – The top-level directory of the dataset. (default: "LibriSpeech")

  • download (bool, optional) – Whether to download the dataset if it is not found at root path. (default: False).

__getitem__(n: int) → Tuple[torch.Tensor, int, str, int, int, int][source]

Load the n-th sample from the dataset.

Parameters

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

Returns

(waveform, sample_rate, utterance, speaker_id, chapter_id, utterance_id)

Return type

tuple

LIBRITTS

class torchaudio.datasets.LIBRITTS(root: str, url: str = 'train-clean-100', folder_in_archive: str = 'LibriTTS', download: bool = False)[source]

Create a Dataset for LibriTTS.

Parameters
  • root (str) – Path to the directory where the dataset is found or downloaded.

  • url (str, optional) – The URL to download the dataset from, or the type of the dataset to dowload. Allowed type values are "dev-clean", "dev-other", "test-clean", "test-other", "train-clean-100", "train-clean-360" and "train-other-500". (default: "train-clean-100")

  • folder_in_archive (str, optional) – The top-level directory of the dataset. (default: "LibriTTS")

  • download (bool, optional) – Whether to download the dataset if it is not found at root path. (default: False).

__getitem__(n: int) → Tuple[torch.Tensor, int, str, str, int, int, str][source]

Load the n-th sample from the dataset.

Parameters

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

Returns

(waveform, sample_rate, original_text, normalized_text, speaker_id, chapter_id, utterance_id)

Return type

tuple

LJSPEECH

class torchaudio.datasets.LJSPEECH(root: str, url: str = 'https://data.keithito.com/data/speech/LJSpeech-1.1.tar.bz2', folder_in_archive: str = 'wavs', download: bool = False)[source]

Create a Dataset for LJSpeech-1.1.

Parameters
  • root (str) – Path to the directory where the dataset is found or downloaded.

  • url (str, optional) – The URL to download the dataset from. (default: "https://data.keithito.com/data/speech/LJSpeech-1.1.tar.bz2")

  • folder_in_archive (str, optional) – The top-level directory of the dataset. (default: "wavs")

  • download (bool, optional) – Whether to download the dataset if it is not found at root path. (default: False).

__getitem__(n: int) → Tuple[torch.Tensor, int, str, str][source]

Load the n-th sample from the dataset.

Parameters

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

Returns

(waveform, sample_rate, transcript, normalized_transcript)

Return type

tuple

SPEECHCOMMANDS

class torchaudio.datasets.SPEECHCOMMANDS(root: str, url: str = 'speech_commands_v0.02', folder_in_archive: str = 'SpeechCommands', download: bool = False)[source]

Create a Dataset for Speech Commands.

Parameters
  • root (str) – Path to the directory where the dataset is found or downloaded.

  • url (str, optional) – The URL to download the dataset from, or the type of the dataset to dowload. Allowed type values are "speech_commands_v0.01" and "speech_commands_v0.02" (default: "speech_commands_v0.02")

  • folder_in_archive (str, optional) – The top-level directory of the dataset. (default: "SpeechCommands")

  • download (bool, optional) – Whether to download the dataset if it is not found at root path. (default: False).

__getitem__(n: int) → Tuple[torch.Tensor, int, str, str, int][source]

Load the n-th sample from the dataset.

Parameters

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

Returns

(waveform, sample_rate, label, speaker_id, utterance_number)

Return type

tuple

TEDLIUM

class torchaudio.datasets.TEDLIUM(root: str, release: str = 'release1', subset: str = None, download: bool = False, audio_ext='.sph')[source]

Create a Dataset for Tedlium. It supports releases 1,2 and 3.

Parameters
  • root (str) – Path to the directory where the dataset is found or downloaded.

  • release (str, optional) – Release version. Allowed values are "release1", "release2" or "release3". (default: "release1").

  • subset (str, optional) – The subset of dataset to use. Valid options are "train", "dev", and "test" for releases 1&2, None for release3. Defaults to "train" or None.

  • download (bool, optional) – Whether to download the dataset if it is not found at root path. (default: False).

__getitem__(n: int) → Tuple[torch.Tensor, int, str, int, int, int][source]

Load the n-th sample from the dataset.

Parameters

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

Returns

(waveform, sample_rate, transcript, talk_id, speaker_id, identifier)

Return type

tuple

property phoneme_dict

Phonemes. Mapping from word to tuple of phonemes. Note that some words have empty phonemes.

Type

dict[str, tuple[str]]

VCTK

class torchaudio.datasets.VCTK(root: str, url: str = 'https://datashare.is.ed.ac.uk/bitstream/handle/10283/3443/VCTK-Corpus-0.92.zip', folder_in_archive: str = 'VCTK-Corpus', download: bool = False, downsample: bool = False, transform: Any = None, target_transform: Any = None)[source]

Create a Dataset for VCTK.

Note

Parameters
  • root (str) – Path to the directory where the dataset is found or downloaded.

  • url (str, optional) – Not used as the dataset is no longer publicly available.

  • folder_in_archive (str, optional) – The top-level directory of the dataset. (default: "VCTK-Corpus")

  • download (bool, optional) – Whether to download the dataset if it is not found at root path. (default: False). Giving download=True will result in error as the dataset is no longer publicly available.

  • downsample (bool, optional) – Not used.

  • transform (callable, optional) – Optional transform applied on waveform. (default: None)

  • target_transform (callable, optional) – Optional transform applied on utterance. (default: None)

__getitem__(n: int) → Tuple[torch.Tensor, int, str, str, str][source]

Load the n-th sample from the dataset.

Parameters

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

Returns

(waveform, sample_rate, utterance, speaker_id, utterance_id)

Return type

tuple

VCTK_092

class torchaudio.datasets.VCTK_092(root: str, mic_id: str = 'mic2', download: bool = False, url: str = 'https://datashare.is.ed.ac.uk/bitstream/handle/10283/3443/VCTK-Corpus-0.92.zip', audio_ext='.flac')[source]

Create VCTK 0.92 Dataset

Parameters
  • root (str) – Root directory where the dataset’s top level directory is found.

  • mic_id (str) – Microphone ID. Either "mic1" or "mic2". (default: "mic2")

  • download (bool, optional) – Whether to download the dataset if it is not found at root path. (default: False).

  • url (str, optional) – The URL to download the dataset from. (default: "https://datashare.is.ed.ac.uk/bitstream/handle/10283/3443/VCTK-Corpus-0.92.zip")

  • audio_ext (str, optional) – Custom audio extension if dataset is converted to non-default audio format.

Note

  • All the speeches from speaker p315 will be skipped due to the lack of the corresponding text files.

  • All the speeches from p280 will be skipped for mic_id="mic2" due to the lack of the audio files.

  • Some of the speeches from speaker p362 will be skipped due to the lack of the audio files.

  • See Also: https://datashare.is.ed.ac.uk/handle/10283/3443

__getitem__(n: int) → Tuple[torch.Tensor, int, str, str, str][source]

Load the n-th sample from the dataset.

Parameters

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

Returns

(waveform, sample_rate, utterance, speaker_id, utterance_id)

Return type

tuple

YESNO

class torchaudio.datasets.YESNO(root: str, url: str = 'http://www.openslr.org/resources/1/waves_yesno.tar.gz', folder_in_archive: str = 'waves_yesno', download: bool = False, transform: Any = None, target_transform: Any = None)[source]

Create a Dataset for YesNo.

Parameters
  • root (str) – Path to the directory where the dataset is found or downloaded.

  • url (str, optional) – The URL to download the dataset from. (default: "http://www.openslr.org/resources/1/waves_yesno.tar.gz")

  • folder_in_archive (str, optional) – The top-level directory of the dataset. (default: "waves_yesno")

  • download (bool, optional) – Whether to download the dataset if it is not found at root path. (default: False).

  • transform (callable, optional) – Optional transform applied on waveform. (default: None)

  • target_transform (callable, optional) – Optional transform applied on utterance. (default: None)

__getitem__(n: int) → Tuple[torch.Tensor, int, List[int]][source]

Load the n-th sample from the dataset.

Parameters

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

Returns

(waveform, sample_rate, labels)

Return type

tuple

Docs

Access comprehensive developer documentation for PyTorch

View Docs

Tutorials

Get in-depth tutorials for beginners and advanced developers

View Tutorials

Resources

Find development resources and get your questions answered

View Resources