SR-CEM: New ASR confidence module cuts calibration error 4.5x

A common problem in automatic speech recognition (ASR) is knowing when the model is wrong. Traditional softmax confidence scores are poorly calibrated, especially for end-to-end systems. To fix this, researchers from KU Leuven propose the Score-Rank Confidence Estimation Module (SR-CEM), a lightweight add-on that extracts features directly from beam search — the decoding process ASR models already use. Instead of relying on internal model architecture, SR-CEM combines the scores and ranks of token hypotheses to produce token- and word-level confidence estimates.

On in-domain English test data, SR-CEM achieves a Maximum Calibration Error of 4.50% and Expected Calibration Error of 0.30% at token level, compared to 20.04% and 1.75% for softmax confidence — a 4.5x improvement in MCE. Word-level results show 8.17% MCE versus 17.91%. Crucially, the method works unchanged across hybrid, transducer, and attention-based architectures with different decoding strategies, and remains robust on Dutch, noisy, and conversational speech. The paper, published in IEEE Open Journal of Signal Processing and presented at ICASSP 2026, positions SR-CEM as a general-purpose solution for producing reliable ASR confidence scores, which is critical for downstream applications like voice assistants, automated subtitling, and human-in-the-loop review systems.