Using multiple acoustic feature sets for speech recognition

In this paper, the use of multiple acoustic feature sets for speech recognition is investigated. The combination of both auditory as well as articulatory motivated features is considered. In addition to a voicing feature, we introduce a recently developed articulatory motivated feature, the spectrum derivative feature. Features are combined both directly using linear discriminant analysis (LDA) as well as indirectly on model level using discriminative model combination (DMC). Experimental results are presented for both small- and large-vocabulary tasks. The results show that the accuracy of automatic speech recognition systems can be significantly improved by the combination of auditory and articulatory motivated features. The word error rate is reduced from 1.8% to 1.5% on the SieTill task for German digit string recognition. Consistent improvements in word error rate have been obtained on two large-vocabulary corpora. The word error rate is reduced from 19.1% to 18.4% on the VerbMobil II corpus, a German large-vocabulary conversational speech task, and from 14.1% to 13.5% on the British English part of the European parliament plenary sessions (EPPS) task from the 2005 TC-STAR ASR evaluation campaign.