[en] This paper describes a new approach for robust speech recognition. Inspired from mask theory, the Frame Dropping technique consists in using only highly voiced frames in the decoding process. The idea is to be more robust to noise by using more stable parts. This method was assessed on a database of isolated words in additive noise provided by the Signal Processing Information Base (NOISEX-92). The test showed that there is a deterioration in clean environment. However, the results are quite close to the baseline in adverse conditions. Moreover, it paves the way for global reduction of complexity.
