An Improved ASR Front-End for Exceeding the AFE Performance on Aurora 3 Task

Inhalt:
We focus on the front-end development for low-resource mobile devices evaluated on small vocabulary recognition task in car noise. We propose the least-squares based weighting rules in the frequency domain using a three-state voice activity andminimumstatistics noise estimators. The weighting rules are combined with the root cepstral coefficients and the cepstral smoothing to further increase the robustness. Recognition results are presented on the Aurora 3 German digits database. The proposed least-squares weighting rules are better than the noise reduction scheme in the ETSI standardized advanced front-end (AFE) and the combined techniques outperform the AFE on the given task.