Fusion of multiple uncertainty estimators and propagators for noise robust ASR

摘要: Uncertainty decoding has been successfully used for speech recognition in highly nonstationary noise environments. Yet, accurate estimation of the uncertainty on denoised signals and propagation to features remain difficult. In this work, we propose fuse estimates obtained from different estimators propagators by linear combination. The fusion coefficients are optimized minimizing a measure divergence with oracle development data. Using Kullback-Leibler divergence, obtain 18% relative error rate reduction 2nd CHiME Challenge respect conventional decoding, that is about twice as much achieved best single estimator propagator.