The study is part of a series investigating production and perception of lexical stress in a number of languages including Brazilian Portuguese, English, Estonian, French, Italian and Swedish. The production database contains data representing male and female speakers in the above languages in three speaking styles – spontaneous speech, phrase reading, and wordlist reading. Keywords from these recordings, representing male and female speakers and all speaking styles are used. The participants’ task is to judge the relative syllable prominences of the keywords presented one by one. In a previous study, subjects were native Swedish speakers. In the present study subjects are native speakers of Italian. In the analyses, perception results are correlated with acoustic variables shown to be important in the production studies. From the previous perception study we know that acoustic syllable prominence affects perceived syllable prominence. But there is also a possibility that listeners’ perception may be biased by expectations based on the listeners’ native language. The main result is that there are great similarities between the Swedish and Italian listeners in the way acoustic prominence affects perceived prominence, but we are also able to demonstrate a case of native language bias.

The present study is the first in a series of studies exploring the perception of lexical stress in a number of languages. As stimuli, key words extracted from recordings in Brazilian Portuguese, English, Estonian, French, Italian and Swedish are used. The data represent male and female speakers in all languages and three different speaking styles – spontaneous speech, phrase reading, and wordlist reading. The ultimate goal of the perception studies is to explore the perception of prominence as a function of the acoustic properties of the stimuli and the native language of the listeners. In this paper we compare the prominence scores assigned to syllables by 44 native Swedish speakers with two automatic methods: acoustic feature analysis using acoustic properties of syllables and continuous wavelet transform. Both methods use duration, F0 and spectral emphasis characteristics of speech signal or a subset thereof. Our results demonstrate a strong language dependency of the way acoustic characteristics correlate with prominence. Correlations between prominence scores and phonological word stress patterns show that the human raters resolve this language-dependency better than the automatic signal-based methods. Also, the signal feature combinations for which the raters’ judgements correlate best with the automatically assigned prominence scores depend on stimulus language to a larger extent that on the signal-based method used.