Utterance-Level Methods for Identifying Reliable ASR-Output for Child Speech

A research team from Radboud University has developed a breakthrough method for improving Automatic Speech Recognition (ASR) accuracy for child speech applications. Their paper, "Utterance-Level Methods for Identifying Reliable ASR-Output for Child Speech," addresses a critical limitation in educational technology: current ASR systems struggle with children's speech patterns, leading to high error rates that undermine language learning and literacy tools. The researchers created two distinct approaches—one for read speech and one for dialogue material—that can identify which transcriptions are reliable before they're used in applications.

The system was evaluated on both English and Dutch datasets using baseline and fine-tuned models, demonstrating remarkable consistency across languages. The best-performing strategy achieved precision scores exceeding 97.4% for both speech types, meaning when the system flags a transcription as reliable, it's correct nearly every time. This allows 21.0% to 55.9% of dialogue and read speech datasets to be automatically selected with minimal error rates (UER <2.6%). The method represents a practical solution that doesn't require perfect ASR, but rather smart filtering of what's already working well.

This research, submitted for Interspeech 2026, provides a crucial bridge between current ASR limitations and practical educational applications. By reliably identifying which child speech recordings have accurate transcriptions, developers can build more effective language learning tools without waiting for perfect speech recognition technology. The approach is particularly valuable for literacy acquisition programs and language learning applications where accurate feedback depends on reliable transcription.