Dataset Card for English MLS

Dataset Summary

This is a streamable version of the English version of the Multilingual LibriSpeech (MLS) dataset. The data archives were restructured from the original ones from OpenSLR to make it easier to stream.

MLS dataset is a large multilingual corpus suitable for speech research. The dataset is derived from read audiobooks from LibriVox and consists of 8 languages - English, German, Dutch, Spanish, French, Italian, Portuguese, Polish. It includes about 44.5K hours of English and a total of about 6K hours for other languages.

This dataset card includes the 44.5K hours of English. Refers to this dataset card for the other languages.

Supported Tasks and Leaderboards

automatic-speech-recognition, speaker-identification: The dataset can be used to train a model for Automatic Speech Recognition (ASR). The model is presented with an audio file and asked to transcribe the audio file to written text. The most common evaluation metric is the word error rate (WER). The task has an active leaderboard which can be found at https://paperswithcode.com/dataset/multilingual-librispeech and ranks models based on their WER.
text-to-speech, text-to-audio: The dataset can also be used to train a model for Text-To-Speech (TTS).

How to use

The datasets library allows you to load and pre-process your dataset in pure Python, at scale. The dataset can be downloaded and prepared in one call to your local drive by using the load_dataset function.

For example, to download the German config, simply specify the corresponding language config name (i.e., "german" for German):

from datasets import load_dataset

mls = load_dataset("parler-tts/mls_eng", split="train")

Using the datasets library, you can also stream the dataset on-the-fly by adding a streaming=True argument to the load_dataset function call. Loading a dataset in streaming mode loads individual samples of the dataset at a time, rather than downloading the entire dataset to disk.

from datasets import load_dataset

mls = load_dataset("parler-tts/mls_eng", split="train", streaming=True)

print(next(iter(mls)))

Bonus: create a PyTorch dataloader directly with your own datasets (local/streamed).

Local:

from datasets import load_dataset
from torch.utils.data.sampler import BatchSampler, RandomSampler

mls = load_dataset("parler-tts/mls_eng", split="train")
batch_sampler = BatchSampler(RandomSampler(mls), batch_size=32, drop_last=False)
dataloader = DataLoader(mls, batch_sampler=batch_sampler)

Streaming:

from datasets import load_dataset
from torch.utils.data import DataLoader

mls = load_dataset("parler-tts/mls_eng", split="train", streaming=True)
dataloader = DataLoader(mls, batch_size=32)

To find out more about loading and preparing audio datasets, head over to hf.co/blog/audio-datasets.

Example scripts

Train your own CTC or Seq2Seq Automatic Speech Recognition models on MultiLingual Librispeech with transformers - here.

Dataset Structure

Data Fields

file: A filename .flac format.
audio: A dictionary containing the audio filename, the decoded audio array, and the sampling rate. Note that when accessing the audio column: dataset[0]["audio"] the audio file is automatically decoded and resampled to dataset.features["audio"].sampling_rate. Decoding and resampling of a large number of audio files might take a significant amount of time. Thus it is important to first query the sample index before the "audio" column, i.e. dataset[0]["audio"] should always be preferred over dataset["audio"][0].
text: the transcription of the audio file.
id: unique id of the data sample.
speaker_id: unique id of the speaker. The same speaker id can be found for multiple data samples.
chapter_id: id of the audiobook chapter which includes the transcription.

Dataset Creation

Curation Rationale

[Needs More Information]

Source Data

Initial Data Collection and Normalization

[Needs More Information]

Who are the source language producers?

[Needs More Information]

Annotations

Annotation process

[Needs More Information]

Who are the annotators?

[Needs More Information]

Personal and Sensitive Information

The dataset consists of people who have donated their voice online. You agree to not attempt to determine the identity of speakers in this dataset.

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

[Needs More Information]

Additional Information

Dataset Curators

[Needs More Information]

Licensing Information

Public Domain, Creative Commons Attribution 4.0 International Public License (CC-BY-4.0)

Citation Information

@article{Pratap2020MLSAL,
  title={MLS: A Large-Scale Multilingual Dataset for Speech Research},
  author={Vineel Pratap and Qiantong Xu and Anuroop Sriram and Gabriel Synnaeve and Ronan Collobert},
  journal={ArXiv},
  year={2020},
  volume={abs/2012.03411}
}

Data Statistics

Duration (h)	Train	Dev	Test
English	44,659.74	15.75	15.55
German	1,966.51	14.28	14.29
Dutch	1,554.24	12.76	12.76
French	1,076.58	10.07	10.07
Spanish	917.68	9.99	10
Italian	247.38	5.18	5.27
Portuguese	160.96	3.64	3.74
Polish	103.65	2.08	2.14

# Speakers	Train		Dev		Test
Gender	M	F	M	F	M	F
English	2742	2748	21	21	21	21
German	81	95	15	15	15	15
Dutch	9	31	3	3	3	3
French	62	80	9	9	9	9
Spanish	36	50	10	10	10	10
Italian	22	43	5	5	5	5
Portuguese	26	16	5	5	5	5
Polish	6	5	2	2	2	2

# Hours / Gender	Dev		Test
Gender	M	F	M	F
English	7.76	7.99	7.62	7.93
German	7.06	7.22	7	7.29
Dutch	6.44	6.32	6.72	6.04
French	5.13	4.94	5.04	5.02
Spanish	4.91	5.08	4.78	5.23
Italian	2.5	2.68	2.38	2.9
Portuguese	1.84	1.81	1.83	1.9
Polish	1.12	0.95	1.09	1.05