Prosodic variants of syllable templates for speech synthesis

In many speech synthesis sytems based on concatenation of prerecorded units, the same segmental template is used in different prosodic environments, typically with simple manipulations of acoustic parameters. The present experiment investigated how successful simple techniques are for syllable-based synthesis. English speakers produced CVC and CV syllables with /p,b,w/ and /i,a,ay/, in isolation and in polysyllabic reiterant phrases (e.g., /bayb bayb bayb bayb/) with different stress and intonation patterns. Pitch tracking the LPC analysis were performed, and syllable templates were excised. Hybrid phrases were created by replacing a syllable template in one prosodic environment with a template for the same syllable from another prosodic environment, substituting the appropriate duration, fundamental frequency, and amplitude. Listening tests showed that natural-sounding hybrid unstressed syllables can be constructed from templates from stressed original syllables, but not vice versa. Also, hybrids should retain the syllable position of the original templates (e.g., phrase-initial, phrase-final). These results are useful for determining which manipulations produce natural-sounding speech, and deciding whether multiple templates for the same segmental unit are necessary.