Rank: 2nd out of 430, winning $15,000
Developing Advanced ASR Algorithms to Understand the Pitch, Rhythm, and Evolving Articulation of Young Learners
About the Competition
The “On Top of Pasketti” challenge was a code execution competition hosted by DrivenData aimed at advancing Automatic Speech Recognition (ASR) technology for children. The competition featured two distinct tracks: a Word Track, where participants predicted exact words from audio clips, and a Phonetic Track, which required predicting speech sounds using the International Phonetic Alphabet (IPA). With a total prize pool of $120,000, participants had to package their models for containerized execution on a single NVIDIA A100 GPU within a strict 2 hours inference time limit. Team Epoch came second on the Phonetic Track winning $15,000, Epoch’s biggest prize win yet.
“Transforming early literacy through innovative teaching tools unlocks children’s potential.”
Gates Foundation
The competition was hosted by DrivenData and sponsored by the Bill & Melinda Gates Foundation. DrivenData is a nonprofit organization that leverages social impact-focused machine learning challenges to bring innovative data science solutions to critical global problems.
Relevance
Current ASR systems perform poorly with children because they are primarily trained on adult voices, failing to account for the unique acoustic and linguistic variations of young learners. Children present challenges such as inconsistent pronunciation, higher fundamental frequencies, and slower speaking rates. Furthermore, speech patterns and articulation shift rapidly as children grow, and there is a significant lack of high-quality, large-scale classroom speech datasets with proper phonetic labels to train more robust models.
Technical Details
The solution utilized a complex ensemble of 13 models using a character-level voting algorithm, ROVER, to achieve a private leaderboard score of 0.2607. The technical approach integrated diverse state-of-the-art architectures including WavLM, Whisper, and HuBERT, optimized with Focal CTC loss and differential learning rates. To address child-specific speech variability, the team employed Minimum Bayes Risk (MBR) decoding and extensive augmentations such as pitch shifting, speed perturbation, and frequency masking.
See our solution on GitHub here:
UN Sustainable Development Goals
This competition directly supports UN Sustainable Development Goal 4: Quality Education by fostering more inclusive and effective educational tools. By improving ASR for children, this work enables the early detection of learning difficulties like dyslexia and supports the creation of interactive learning applications tailored to young learners. Additionally, the initiative promotes fairness in AI by addressing data gaps in underrepresented dialects and real-world classroom environments.
