Great expectations - introspective vs. perceptual prominence ratings and their acoustic correlates

In order to gain knowledge about the interaction between topdown expectations of listeners concerning prosodic prominence and its acoustic correlates, two exploratory empirical studies were carried out. First, native and nonnative subjects rated prominences of speech read at normal and very fast —prosodically very different — speech. Later, these ratings were compared with introspective prominence ratings of different listeners. First results indicate a major influence of the introspection on prominence ratings, especially if acoustic cues are difficult to interpret, as it is the case in very fast speech. Compared to native subjects, nonnatives rely less on their introspection and more on the acoustics. 1. Acoustic and Phonological Prominence Prediction and Perception Algorithms of prominence prediction that are based on phonological and morpho-syntactic information perform quite well. Their evaluation on the basis of perceptual prominence ratings shows good results (e.g. [1]). Also, the acoustic correlates of perceptual prominence are quite well-known – F0, duration and spectral tilt have been claimed to be the most important cues for a number of languages, even though the relative proportion of each parameter may be language specific. It is also well-known that the listeners’ psychoacoustic sensitivity for variations in duration, F0 or intensity differs dramatically between speech and non-speech stimuli [2]. One of the reasons for this is that prominence and its verified acoustic correlates do not stand in a 1:1 relationship. Instead, there is a complex interaction between acoustic prosody and listeners’ top down expectancies concerning lexical or sentence stress. The latter is based on their linguistic competence, the former based on their perception. [3] detected high correlations between perceptual prominence ratings and well-known acoustic correlates. But they claim that: 1 Spectral tilt or spectral slope is commonly regarded as the relative intensity of high versus low frequencies. No standardized measurement of spectral tilt has been established yet. “The results show that subjects can use vocal effort, the distinctness of F0-movements, and vowel duration as cues for rating syllable prominence. However, we can not tell which cues they actually used. A strategy based mainly on top-down processing could have produced a similar result”. It is still an unresolved question, to what extent perceived prominence is independent of top-down hypotheses based on the speaker’s native language competence. In other words, we do not know whether a lexical stress is perceived because it has been produced or because the speaker expects it to be produced. This lack of knowledge may be an explanation, why it has been difficult finding clear evidence for well-established phonological concepts such as stress shift [4]. Also, it has been shown that the native language makes speakers more or less sensitive to the acoustic cues of prominence [5]. E.g., native speakers of a language with fixed lexical stress may have problems detecting it in a language with free lexical stress [6]. The strong influence of top-down expectations concerning prominence perception receives further support by research from speech synthesis. Here, it was often found that prominence prediction based on linguistic knowledge performs better than an acoustic prediction [7, 8]. The strong influence of top-down knowledge would also explain the high inter speaker and inter-listener correlation of prominence patterns that was sometimes found (e.g. [9]). It would be of course nonsense to assume an independence of prominence perception and acoustic correlates. If this were the case listeners would be unable to detect deviances of the “expected” prominence pattern, as it happens frequently in contrastive utterances, corrections or faulty and consequently often hard to understand L2productions. Also, without a clear correspondence between acoustic prominence and phonological words, listeners would not have a chance to train their top-down expectancies concerning prominence during the process of language acquisition. 2. Research Questions and Hypotheses Prominence perception cannot be explained on the basis of either acoustic prosody or top-down expectancies alone. Thus, there must be a trade-off between both factors influencing prominence perception. It is likely, that listeners possess a certain tolerance concerning speaker specific or speaking style specific variations. For example, a native listener may extrapolate certain expected pitch excursions even if the pitch contour he listens to remains rather flat. However, strong deviations from an expected acoustic correlate must be perceptible since they may have an important communicative function or signal. An example for September, 4-8, Lisbon, Portugal INTERSPEECH 2005