Boundary-Robust Lightweight Speech Command Recognition via Adaptive Noise-Floor Estimation and Reliability-Conditioned Cross-Scale Fusion

Abstract

This paper presents a boundary-robust lightweight speech-command recognizer based on adaptive quantile noise-floor estimation (AQNE) and reliability-conditioned cross-scale interaction fusion (RCIF). AQNE estimates the noise floor from the lowest-energy 20% of overlapping frames and provides both SNR and reliability cues. RCIF enables three temporal branches to exchange a shared residual before conservative quality-conditioned fusion. During training, a shifted noisy student view is regularized by a clean centered exponential-moving-average (EMA) teacher, which is removed at inference. On a verified 1,500-recording spoken-digit dataset with white and pink noise, the revised model achieved 98.11±0.19% accuracy at 0 dB and 95.22±0.84% at −5 dB across three seeds. At a target SNR of 10 dB, AQNE’s error remained 0.84 dB under ±300-ms displacement, whereas the fixed-boundary estimator reached 8.44 dB. At −5 dB and ±300 ms, accuracy improved from 93.50% to 94.50%. The model contains 147.7K parameters, occupies 0.591 MB in FP32, and requires 0.871 ms per command on the evaluation CPU.