CHILDES-Aligned dataset boosts ASR for kids with 19.5% WER reduction

A major hurdle for child speech recognition is the lack of clean, accurately timed audio data. The CHILDES corpus, while rich in naturalistic child-adult interactions, suffers from utterance-level timestamps that are often noisy, incomplete, or misaligned with the audio. Researchers from UIUC, IBM Research, and NTU Singapore have now introduced BEACON (Boundary Estimation via Alignment CONsensus), a framework that cleans up these timestamps by pooling predictions from multiple off-the-shelf ASR models. Each model generates word-level timestamps aligned to the provided human transcripts, and a consensus voting strategy determines the final utterance boundaries. The approach is corpus-agnostic, meaning it can be applied to any long-form recording paired with a trusted transcript but unreliable timestamps.

Using this pipeline, the team curated and released CHILDES-Aligned, a 413-hour general-purpose child-speech dataset with corrected timestamps. From that, they extracted a 283-hour quality-controlled subset specifically designed for ASR training. Fine-tuning an ASR model on this subset produced an average 19.5% relative Word Error Rate (WER) reduction across four out-of-domain child-speech benchmarks. This represents a significant leap for child speech processing, an area historically limited by data quality. The dataset and BEACON framework are both publicly available, offering a scalable recipe for timestamp curation that could accelerate research in child language acquisition, educational tools, and voice interfaces for children.