On Top of Pasketti: Children’s Speech Recognition Challenge - Word Track

Develop automatic speech recognition models that produce word-level transcriptions of children’s speech. #education

$70,000 in prizes
6 weeks left
346 joined

Overview

Voice is one of the most natural ways for kids to learn, explore, and show what they know, yet today's Automatic Speech Recognition (ASR) technology hardly understands them. Built on adult speech, most ASR systems struggle with the pitch, rhythm, and evolving articulation of young learners.

The high error rate of ASR for children prevents downstream uses that could enhance educational outcomes and scale early screening and intervention. High-quality ASR for kids can unlock a new generation of learning tools like playful voice-driven tutoring systems, early support for speech and phonological development, adaptive literacy assessments, and accessible interfaces for students who are preliterate or otherwise benefit from alternatives to text or visual input.

Challenge structure

This challenge assembles pre-existing and newly labeled datasets to advance speech models that truly work for children. Solvers will develop models that accurately capture what children say and how they say it from audio recordings.

There are two independent tracks that enable different downstream uses. Solvers can compete in either track and can earn prizes for each. Across tracks, prizes total $120,000.

  • In the Word Track (YOU ARE HERE), solvers will predict the words spoken by children in audio clips. Word-level models enable automated transcription, verbal tool use, and assessments related to cognition and speech (e.g., comprehension, reasoning).
  • In the Phonetic Track, solvers will predict the speech sounds, or phones, spoken by children in audio clips. Phonetic models are critical for diagnostic applications like speech pathology screening.

Prizes

Competition End Date:

April 6, 2026, 11:59 p.m. UTC

Place Prize Amount
1st $25,000
2nd $15,000
3rd $10,000
Noisy Classroom Bonus (4x) $5,000 each

Noisy Classroom Bonus

Strong performance in noisy, real-world classroom environments is critical for educational ASR use cases. Teams that place in the top 20 on the Word Track leaderboard will be eligible for this bonus prize. The four teams with the best performance on the Noisy Word Error Rate (Noisy WER) metric will each receive $5,000.

Update February 12, 2026 - RealClass Data Release: We have released a new dataset with synthetic classroom background noise that can be used to help improve model performance on real-world classroom environments by combining that noise with the other competition datasets.


Competition Rules

The competition rules are designed to promote fair competition and encourage useful, responsible solutions. If you are ever unsure whether your solution meets the competition rules, ask in the competition forum or send an email to info@drivendata.org. Key rules are highlighted below.

See the full competition rules for complete details.


Competition Data

Participants must:

  • Use data only for this challenge and only during the challenge period. After the competition ends, information will be provided about a subset of data sources that can be accessed and used beyond the challenge.

  • Ensure that any competition data stored on a laptop, portable device, or removable storage medium are encrypted (for example, by enabling full-disk encryption such as FileVault on macOS or BitLocker on Windows).

  • Delete all local data after the competition ends, unless a separate license allows continued use.

  • Never share, copy, or publish the data. For example, you may not use tools like Codex and ChatGPT that store or retain uploaded data, though you may download model weights and run models locally.

  • Never try to identify or contact anyone who contributed data to the study.

  • Immediately contact DrivenData if anyone gains unauthorized access to the data or if you inadvertently identify someone.

  • Include appropriate attribution in any publication, presentation, or public description of your competition work. A full list of citations will be posted on the About page after the submission deadline.


External Data and Models

Updated February 13, 2026 to clarify requirements for privately held external data.

External data and pre-trained models are allowed in this competition, except those in or trained on the prohibited sources listed below. Participants may not use tools like Codex or ChatGPT that store or retain uploaded data, but may download model weights and run models locally.

This challenge aims to support open solutions with broad social benefit and real-world applicability. To be eligible for prizes, any external data or pre-trained models used must be licensed so that the resulting model can be released for broad use, in and beyond the competition, including for commercial purposes (no NC, CC NC, or CC BY-NC licenses).

Participants may use external data provided they have the legal right to do so. While this data does not need to be shared publicly, the data must be shareable with the challenge organizers to allow for independent result verification and the development of openly licensed models in order to be eligible for prizes. See the external data section of in the challenge rules for further details.

If you have questions about licensing in general or whether specific external data and models can be used, post in the competition forum or send an email to info@drivendata.org.

Top teams on the leaderboard will be required to declare all external data and pre-trained models used. Each team must either: (1) certify that all resources are licensed to enable commercial model use and provide documentation if requested, or (2) opt out of prize eligibility.

Prohibited Data Sources

Participants may not use any of the following external datasets or data providers, nor any pre-trained models that are known to have been trained on them, except where explicitly provided through the challenge:

  • Arizona Acoustic Child Database
  • CMU Kids Corpus
  • CSLU: Kids' Speech Version 1.1
  • MyST Children's Conversational Speech
  • Reach Every Reader (RER) ReadNet Dataset
  • RealClass Dataset
  • Speech Production Repository to Optimizing Use of AI Technologies (SPROUT) Dataset
  • TalkBank (except for the hosted competition data linked from the Data Download page)
  • TeachFX

Solver Eligibility

This competition is open only to natural persons (no organizations or legal entities may participate).

Any officers, directors, employees, advisory board members, or contractors of DrivenData, as well as their immediate family or household members are not eligible to win a prize. Officers, directors, employees, and advisory board members of competition funders and data providers are also not eligible to win a prize.


How to Compete

  1. Click the "Compete!" button in the sidebar to enroll in the competition.
  2. Get familiar with the problem through the overview and problem description. You might also want to reference additional resources available on the about page.
  3. Download the data from the data download page, and set a reminder to delete it after the competition ends.
  4. Create and train your own model.
  5. Package your model files with the code to make predictions based on the runtime repository specification on the code submission format page.
  6. From the submissions page, submit your code as a zip archive for containerized execution. You're in!

This challenge is funded by the Gates Foundation, with additional funding support from the Valhalla Foundation, TeachFX, and the University of Maryland's Center for Educational Data Science and Innovation.


Image Credit: Image generated using Nano Banana; source images include an image by CDC on Unsplash and an image by tolmacho from Pixabay