Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The People's Speech: A Large-Scale Diverse English Speech Recognition Dataset for Commercial Usage

Daniel Galvez, Greg Diamos, Juan Ciro, Juan Felipe Cerón, Keith Achorn, Anjali Gopi, David Kanter, Maximilian Lam, Mark Mazumder, Vijay Janapa Reddi

TL;DR

The Peoples' Speech presents a large-scale, commercially usable English ASR dataset (CC-BY/CC-BY-SA/public-domain) assembled from web-hosted transcripts and aligned via a GPU-accelerated forced-alignment pipeline. It demonstrates the feasibility of curating diverse, license-compliant speech data at scale and provides an initial evaluation showing meaningful ASR signal on Librispeech. The work also addresses licensing, provenance, and maintenance considerations through MLCommons sponsorship, and discusses limitations and future directions for broader language coverage and ongoing dataset updates.

Abstract

The People's Speech is a free-to-download 30,000-hour and growing supervised conversational English speech recognition dataset licensed for academic and commercial usage under CC-BY-SA (with a CC-BY subset). The data is collected via searching the Internet for appropriately licensed audio data with existing transcriptions. We describe our data collection methodology and release our data collection system under the Apache 2.0 license. We show that a model trained on this dataset achieves a 9.98% word error rate on Librispeech's test-clean test set.Finally, we discuss the legal and ethical issues surrounding the creation of a sizable machine learning corpora and plans for continued maintenance of the project under MLCommons's sponsorship.

The People's Speech: A Large-Scale Diverse English Speech Recognition Dataset for Commercial Usage

TL;DR

The Peoples' Speech presents a large-scale, commercially usable English ASR dataset (CC-BY/CC-BY-SA/public-domain) assembled from web-hosted transcripts and aligned via a GPU-accelerated forced-alignment pipeline. It demonstrates the feasibility of curating diverse, license-compliant speech data at scale and provides an initial evaluation showing meaningful ASR signal on Librispeech. The work also addresses licensing, provenance, and maintenance considerations through MLCommons sponsorship, and discusses limitations and future directions for broader language coverage and ongoing dataset updates.

Abstract

The People's Speech is a free-to-download 30,000-hour and growing supervised conversational English speech recognition dataset licensed for academic and commercial usage under CC-BY-SA (with a CC-BY subset). The data is collected via searching the Internet for appropriately licensed audio data with existing transcriptions. We describe our data collection methodology and release our data collection system under the Apache 2.0 license. We show that a model trained on this dataset achieves a 9.98% word error rate on Librispeech's test-clean test set.Finally, we discuss the legal and ethical issues surrounding the creation of a sizable machine learning corpora and plans for continued maintenance of the project under MLCommons's sponsorship.
Paper Structure (15 sections, 7 figures, 2 tables)

This paper contains 15 sections, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Dataset Description
  4. Licensing Description
  5. Transcript Content Characterization
  6. Audio Characterization
  7. Dataset Construction
  8. Data Acquisition
  9. Forced Alignment
  10. Forced Alignment Verification
  11. System Implementation
  12. Dataset Limitations
  13. Evaluation
  14. Discussion
  15. Conclusions

Figures (7)

  • Figure 1: Hours of audio by license type. The "US Government" works are in the public domain. We separate the category from "Public Domain" to highlight how large it is relative to others.
  • Figure 2: Hours of Audio per Language. The majority of our dataset is comprised of English followed by a wide range of other languages.
  • Figure 3: Number of occurrences of each location in the transcripts. For clarity, only those places that appear in the data at least 2500 times were taken into account. The "USA" entity was removed because it dwarved the other locations in number of appearances.
  • Figure 4: Hours of Audio per Sampling Frequency. Since most speech recognition systems use 8kHz or 16 kHz audio data, our data is suitable for training these models via downsampling.
  • Figure 5: Hours of Each Category. Detected by "Bart large MNLI."
  • ...and 2 more figures