Métodos de interpolação para reconstrução de dados faltantes em sistemas robustos de reconhecimento automático de fala

Abstract

The missing data approach was developed to perform automatic speech recognition in noisy environments. This technique identifies and uses in the recognition process only parts of a noisy utterance signal which were not heavily corrupted by the noise, these parts are called reliable. There are two main methods that can be used to achieve this goal: the marginalization and the imputation. The marginalization method uses only the utterance reliable information, whereas the imputation method tries to substitute the unreliable parts for estimates based on the reliable information. The purpose of this paper is to compare three imputation methods: the linear interpolation, the polynomial interpolation and the rational interpolation.