Challenge

Where the Wild Features Are

Predict kindergarteners' end-of-year reading scores from literacy screening audio, and find which speech features carry signal beyond correctness scores

Tasks
Reading assessment
Data types
Audio
Tabular
$50,000 in prizes
8 weeks left model submissions
53 joined
Adapted from a photo by Guy Basabose on Unsplash

About the project


Literacy screeners give teachers an early indication of which children may struggle to learn to read, but they typically record only whether each answer was right or wrong. The audio of a child's response holds much more: how quickly they answered, how they paused, and how they produced each sound. Your work can show which of these signals help predict reading outcomes. Features that are interpretable and grounded in the audio could make early screening more accurate, point researchers toward new markers of reading difficulty, and help children get support before they fall behind.

Reach Every Reader

Reach Every Reader (RER) is a research initiative dedicated to tackling childhood literacy challenges. RER focuses on developing pre-literacy skills in early childhood, creating diagnostic tools for literacy assessment, supporting early intervention and readers' progress, providing educators with literacy resources, and collaborating with researchers and organizations to scale their solutions. Learn more in the video below.

About the data


The RER literacy screening assessment

The data in this challenge come from the RER literacy screening assessment, collected by researchers at MIT and Florida State University (FSU). Children completed the screener at the beginning of their kindergarten year. It was administered on a tablet by a trained adult, who scored each response in the moment. This challenge includes audio from three tasks:

  • Blending: The child listens to a word that is broken down into parts and is then asked to combine the sounds and say the target word. Measures a child's phonological awareness skills.
  • Deletion: The child listens to a word and is then asked to delete a portion of the word and produce a new word (the target word). Measures a child's phonological awareness skills.
  • Sentence repetition: The child listens to a sentence recorded on the tablet and then repeats the sentence verbatim. Assesses the child's ability to listen to and repeat sentences of varying length and complexity without changing the sentence's structure or meaning.

iPad screenshots showing a blending task and a deletion task

iPad screenshots showing the view of the person administering and scoring the literacy screener.

Students were tested on each task until they stopped answering correctly, so the number of clips varies by student.

End-of-year outcomes

At the end of kindergarten, each child took a standardized letter-word reading assessment. The data come from two collection efforts that used different assessments:

  • KTEA-3 Letter & Word Recognition (Kaufman Test of Educational Achievement, Third Edition)
  • WJ IV Letter-Word Identification (Woodcock-Johnson IV Tests of Achievement)

Both are age-normed standard scores of letter-word reading, and they are combined into a single prediction target. Part of the data was collected during the COVID-19 pandemic, when many students enrolled online.

Audio and anonymization

The recordings reflect real classroom conditions. Clips can include background noise and the voices of teachers or test administrators, and a clip scored correct may not contain the child's answer. Item scores come from the trained administrators and are treated as accurate even when the audio misses the response. Strong solutions will rely on signal in the child's speech rather than on adult speech or room acoustics.

To protect children's privacy, every clip was first denoised with FRCRN and then anonymized with FreeVC24 voice conversion, which replaces all voice in a clip with one of about 50 target voices from an age-matched public child speech dataset. Voice conversion preserves much of what was said but may alter some fine acoustic detail. See the problem description for audio format details.

Previous research


This challenge builds on Goodnight Moon, Hello Early Literacy Screening, an earlier DrivenData challenge with the same research team. Participants built models to score whether children's literacy screener responses were correct, using anonymized audio from the same screener. This challenge asks a different question: given the screening scores, what more does the audio say about a child's reading outcome at the end of the year?

Pretrained speech models are allowed in this challenge and may be a useful starting point for extracting acoustic features:

Acoustic features of children's speech

Correctness scores capture whether a child answered right, but not how they said it. Acoustic measures of children's speech can be computed automatically and may carry information that correctness scores miss. In a study of beginning readers reading aloud, pause statistics, syllable rate, and spectral and intensity dynamics were reliable indicators of specific types of oral reading deficits, and performed close to an approach based on automatic speech recognition (Vitthal et al., 2021). This challenge asks whether signals like these, measured in kindergarten screener audio, add to what the screening scores already say about end-of-year reading.

About the sponsors


This project is a partnership between researchers at MIT and FSU:

  • The Gabrieli Lab at MIT aims to understand how the human brain empowers learning, thinking, and feeling, and to use that understanding to help people live happier and more productive lives.

  • The Senseable Intelligence Group at MIT uses neuroimaging, speech communication, and machine learning to improve assessments and treatments for mental health and neurological disorders.

  • The Florida Center for Reading Research at FSU improves learning and achievement from birth through adulthood through rigorous and robust research, innovation, and engagement.

Additional resources


For background on speech features and early reading:

If you're new to extracting features from speech audio, these are useful starting points:

  • Parselmouth, a Python interface to Praat for measures such as pitch, intensity, and formants
  • librosa, a Python library for audio analysis