Human listeners use lexical stress for word segmentation and disambiguation. We look into using lexical stress for speech recognition by examining a Dutch-language corpus. We propose that different spectral features are needed for different phonemes and that, besides vowels, consonants should be taken into account.
展开▼
