MLCommons/peoples_speech

Dataset Card for People's Speech

Dataset Summary

The People's Speech Dataset is among the world's largest English speech recognition corpus today that is licensed for academic and commercial usage under CC-BY-SA and CC-BY 4.0. It includes 30,000+ hours of transcribed speech in English languages with a diverse set of speakers. This open dataset is large enough to train speech-to-text systems and crucially is available with a permissive license.

Supported Tasks and Leaderboards

[Needs More Information]

Languages

English

Dataset Structure

Data Instances

{ "id": "gov_DOT_uscourts_DOT_scotus_DOT_19-161/gov_DOT_uscourts_DOT_scotus_DOT_19-161_DOT_2020-03-02_DOT_mp3_00002.flac", "audio": { "path": "gov_DOT_uscourts_DOT_scotus_DOT_19-161/gov_DOT_uscourts_DOT_scotus_DOT_19-161_DOT_2020-03-02_DOT_mp3_00002.flac" "array": array([-6.10351562e-05, ...]), "sampling_rate": 16000 } "duration_ms": 14490, "text": "contends that the suspension clause requires a [...]" }

Data Fields

{ "id": datasets.Value("string"), "audio": datasets.Audio(sampling_rate=16_000), "duration_ms": datasets.Value("int32"), "text": datasets.Value("string"), }

Data Splits

We provide the following configurations for the dataset: cc-by-clean ("clean"), cc-by-dirty ("dirty"), cc-by-sa-clean ("clean_sa"), cc-by-sa-dirty ("dirty_sa"), and microset ("microset"). We also provide validation and test configurations, which are not only available as standalone configurations but are also included as validation and test splits within each of the above configurations for ease of use. Specifically:

• Setting data_dir="validation" and split="validation" corresponds to the validation split of any of the configurations: "clean", "clean_sa", "dirty", or "dirty_sa".
• Similarly, setting data_dir="test" and split="test" corresponds to the test split of these configurations.

├── clean
│   ├── train
│   ├── validation
│   └── test
├── clean_sa
│   ├── train
│   ├── validation
│   └── test
├── dirty
│   ├── train
│   ├── validation
│   └── test
├── dirty_sa
│   ├── train
│   ├── validation
│   └── test
├── microset
│   └── train
├── validation
│   └── validation
└── test
    └── test

Dataset Creation

Curation Rationale

See our paper.

Source Data

Initial Data Collection and Normalization

Data was downloaded via the archive.org API. No data inference was done.

Who are the source language producers?

[Needs More Information]

Annotations

Annotation process

No manual annotation is done. We download only source audio with already existing transcripts.

Who are the annotators?

For the test and dev sets, we paid native American English speakers to do transcriptions. We do not know the identities of the transcriptionists for data in the training set. For the training set, we have noticed that some transcriptions are likely to be the output of automatic speech recognition systems.

Personal and Sensitive Information

Several of our sources are legal and government proceedings, spoken histories, speeches, and so on. Given that these were intended as public documents and licensed as such, it is natural that the involved individuals are aware of this.

Considerations for Using the Data

Social Impact of Dataset

The dataset could be used for speech synthesis. However, this requires careful cleaning of the dataset, as background noise is not tolerable for speech synthesis.

The dataset could be used for keyword spotting tasks as well. In particular, this is good use case for the non-English audio in the dataset.

Our sincere hope is that the large breadth of sources our dataset incorporates reduces existing quality of service issues today, like speech recognition system’s poor understanding of non-native English accents. We cannot think of any unfair treatment that come from using this dataset at this time.

Discussion of Biases

Our data is downloaded from archive.org. As such, the data is biased towards whatever users decide to upload there.

Almost all of our data is American accented English.

Other Known Limitations

As of version 1.0, a portion of data in the training, test, and dev sets is poorly aligned. Specifically, some words appear in the transcript, but not the audio, or some words appear in the audio, but not the transcript. We are working on it.

Additional Information

Dataset Curators

[Needs More Information]

Licensing Information

We provide CC-BY and CC-BY-SA subsets of the dataset.

Citation Information

Please cite:

@article{DBLP:journals/corr/abs-2111-09344,
  author    = {Daniel Galvez and
               Greg Diamos and
               Juan Ciro and
               Juan Felipe Cer{\'{o}}n and
               Keith Achorn and
               Anjali Gopi and
               David Kanter and
               Maximilian Lam and
               Mark Mazumder and
               Vijay Janapa Reddi},
  title     = {The People's Speech: {A} Large-Scale Diverse English Speech Recognition
               Dataset for Commercial Usage},
  journal   = {CoRR},
  volume    = {abs/2111.09344},
  year      = {2021},
  url       = {https://arxiv.org/abs/2111.09344},
  eprinttype = {arXiv},
  eprint    = {2111.09344},
  timestamp = {Mon, 22 Nov 2021 16:44:07 +0100},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2111-09344.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}