A combination of speaker normalization and speech rate normalization for automatic speech recognition

Abstract

In this contribution a normalization procedure for automatic speech recognition is introduced which aims at reducing speaking rate specific variations of the features of the phonetic classes. A “spurtwise” calculation of normalization factors allows to capture changes of the speaking rate within one utterance. The costsaving implementation using linear interpolation of the original features and a word graph rescoring procedure leads to a moderate increase in computational load compared to the baseline system without speech rate normalization. In addition a two-step procedure which combines vocal tract length normalization (VTLN) and speech rate normalization (SRN) has been developed. Experiments showed, that applying SRN to a VTLN-based recognition system leads to relative reduction in word error rate of 4.2%. This is comparable to the decrease observed when using SRN on a system without VTLN. All in all the combination of VTLN and SRN results in a 15% reduction of word error rate compared to the baseline system.