Martine Grice

The SUS test: a method for the assessment of text-to-speech synthesis intelligibility using semantically unpredictable sentences

This paper describes the experimental set-up used by the SAM (ESPRIT-BRA Project no. 2589: Multilingual Speech Input/Output: Assessment, Methodology and Standardisation) group for evaluating the intelligibility of text-to-speech systems at sentence level. The SUS test measures overall intelligibility of Semantically Unpredictable Sentences which can be automatically generated using five basic syntactic structures and a number of lexicons containing the most frequently occurring mini-syllabic words in each language. The sentence material has the advantage of not being fixed, as words can be extracted from the lexicons randomly to form a new set of sentences each time the test is run. Various text-to-speech systems in a number of languages have been evaluated using this test. Results have demonstrated that the SUS test is effective and that it allows for reliable comparison across synthesisers provided guidelines are followed carefully regarding the definition of the test material and actual running of the test. These recommendations are the result of experience gained during the SAM project and beyond. They are presented here so as to provide users with a standardized evaluation method which is flexible and easy to use and is applicable to a number of different languages.

On the Place of "Phrase Accents" in Intonational Phonology

Many theories of intonational phonology have granted some special status to pitch features that occur at the edges of prosodic domains, contrasting them with prominence-lending pitch configurations. The standard American structuralist theory that flourished in the 1950s (Trager & Smith 1951) drew a clear distinction between PITCH PHONEMES and JUNCTURE PHONEMES , the former constituting the body of a contour and the latter describing the movements at the contour’s end. Parallel to this development, a distinction was also drawn within the Prague School between the cumulative and delimitative functions of tonal phenomena (Trubetzkoy 1958), the former including prominence, the latter domainedge marking. Bolinger (especially 1970) distinguished ‘accent’ from ‘intonation’: ACCENT referred to the distinctive pitch shapes that accompany prominent stressed syllables (now generally known, following Bolinger, as pitch accents), while INTONATION included, among other things, distinctive pitch movements at the ends of contours. A distinction very similar, but not identical, to Bolinger’s is made in the theory of intonation developed at the Institute for Perception Research (IPO) in the Netherlands (Cohen & ’t Hart 1967, ’t Hart et al. 1990), namely between PROMINENCE-LENDING and NON-PROMINENCE-LENDING pitch movements.

The intonation of interrogation in Palermo Italian : implications for intonation theory

In Palermo Italian yes-no interrogatives, if the last syllable of a phrase is unstressed, the nuclear pitch contour is rising-falling, whereas if it is stressed, the contour is simply rising. An autosegmental analysis, where pitch configurations are expressed in terms of H(igh) and L(ow) tones which are either part of a pitch accent or at intermediate or intonation phrase boundaries, is shown to offer the flexibility necessary to account for such context-dependent variation. The interrogative marker consists of a L*+H pitch accent. There is no paradigmatic contrast on the intermediate phrase boundary tone (it is always L) which means that its function is purely delimitative. This tone is only fully realised when a postaccentual syllable is available to carry it; technically, it requires a secondary attachment to a syllable. The absence of the falling part of the L*+HL (L) configuration in phrases with no postaccentual syllable is thus explained.

Leading tones and downstep in English

In early autosegmental studies certain segments were specially marked with a ‘star’ diacritic. These starred segments had a priority association; only in a later stage of the derivation were other segments associated. Starred associations were resistant to modification by subsequent rules: ‘the point is to preserve the prominence of the star’ (Goldsmith 1979: 129). An associated tone had as its phonetic exponent a high or low pitch occurring approximately at the same time as the associated item(s) on the phoneme tier, usually a vowel.

Consistency in transcription and labelling of German intonation with GToBI

A diverse set of speech data was labelled in three sites by 13 transcribers with differing levels of expertise, using GToBI, a consensus transcription system for German intonation. Overall inter-transcriber-consistency suggests that, with training, labellers can acquire sufficient skill with GToBI for large-scale database labelling.

Sources of variation in tonal alignment: Evidence from acoustic and kinematic data

Abstract This study is concerned with the alignment of f0 peaks in rising LH pitch accents in German, both in relation to acoustically defined segments, referred to as segmental anchors, as well as to dynamically defined speech gestures, referred to as articulatory anchors. The effects investigated were the effects of syllable structure (test words ˈCV:CV and ˈCVCV, where the test syllable is open or closed, respectively), dialectal background (the varieties of German spoken in Dusseldorf and Vienna), and accent status in the intonational hierarchy (prenuclear and nuclear accents). As reported for related languages, peaks in closed syllables tended to be later than those in open syllables. However, it was only in nuclear accents that those differences were systematic for all four speakers. Thus only limited support can be provided for an alignment with the syllable edge. Although there was a tendency for Southern varieties to have later peaks than Northern ones, as also found in previous studies, alignment latencies of individual speakers in the two dialectal groups overlapped. These results support a gradient view of dialectal variation in tonal alignment. In this view, dialectal differences are not represented symbolically. Rather, the rising accents used by speakers of both varieties can be adequately captured with one symbolic representation. When comparing prenuclear and nuclear accents, by contrast, differences were found which could be interpreted as discrete. Whereas nuclear accent peaks were anchored to the intervocalic consonant, prenuclear accent peaks were anchored to the following unstressed vowel. This anchor shift could clearly be observed both in the acoustic and articulatory records, reflecting a difference at the symbolic level, possibly in terms of an additional tone following the LH complex.

Assessing incomplete neutralization of final devoicing in German

Abstract It has been claimed that the long established neutralization of the voicing distinction in domain final position in German is phonetically incomplete. However, many studies that have advanced this claim have subsequently been criticized on methodological grounds, calling incomplete neutralization into question. In three production experiments and one perception experiment we address these methodological criticisms. In the first production study, we address the role of orthography. In a large scale auditory task using pseudowords, we confirm that neutralization is indeed incomplete and suggest that previous null results may simply be due to lack of statistical power. In two follow-up production studies (Experiments 2 and 3), we rule out a potential confound of Experiment 1, namely that the effect might be due to accommodation to the presented auditory stimuli, by manipulating the duration of the preceding vowel. While the between-items design (Experiment 2) replicated the findings of Experiment 1, the between-subjects version (Experiment 3) failed to find a statistically significant incomplete neutralization effect, although we found numerical tendencies in the expected direction. Finally, in a perception study (Experiment 4), we demonstrate that the subphonemic differences between final voiceless and “devoiced” stops are audible, but only barely so. Even though the present findings provide evidence for the robustness of incomplete neutralization in German, the small effect sizes highlight the challenges of investigating this phenomenon. We argue that without necessarily postulating functional relevance, incomplete neutralization can be accounted for by recent models of lexical organization.

The Perception of Negative Bias in Bari Italian Questions

This paper investigates two types of questions. OBJECTS, challenging the interlocutor’s assumption that information is shared, are biased towards a negative answer, whereas QUERIES, asking for new information, are generally neutral. In Bari Italian they are both produced with the same pitch accent. However, the height of the pitch peak tends to be greater in OBJECTS. To investigate the perceptual relevance of peak height in distinguishing between these two question types, we carried out a semantically motivated identification task, followed by a discrimination task, recording reaction times in both cases. Results show that listeners can categorically interpret utterances as QUERY or OBJECT on the basis of peak height only. However, their ability to discriminate between pairs of stimuli is poor, thus providing further evidence that categorical interpretation of intonation (like that of vowels) is possible in a labelling task, where listeners have to access linguistic knowledge, whereas it is not possible in a discrimination task, where listeners rely predominantly on psychoacoustic abilities. Results point to the necessity for including [high peak] in the phonology of this variety.

More than a magic moment – Paving the way for dynamics of articulation and prosodic structure

Abstract Research into human communication through the spoken language is full of dichotomies that have often stood in the way of progress in the past, notably the distinction between phonetics and phonology, and more recently, and somewhat orthogonally, between prosody and articulation. The papers collected here make considerable advances in overcoming these restrictions, providing valuable contributions towards the integration of these fields. The increasing evidence for dependencies across the different levels of linguistic structure, and the complexity of the interplay between them, has led to the application of dynamical approaches to spoken language description. With these approaches, coordination and variation within and across systems have begun to play a central role. This paper identifies a common thread through the papers in this issue, in which variation is a consequence of dynamically time-varying behavior that cannot be captured by static snapshots (magic moments).

The effect of focus marking on supralaryngeal articulation – Is it mediated by accentuation?

Abstract In this study we explore the effects of focus-background structure on accentuation (i.e. whether a word bears a pitch accent or not) and supralaryngeal articulation, measured in terms of acoustic durations (syllable and foot durations) and lip kinematics (parameters relating to the opening gesture: duration, displacement, peak velocity and stiffness). Although words in focus were accented and those out of focus were not, there were few supralaryngeal differences between accented words when they were produced in the broad focus context and unaccented target words (out of focus). Thus, accentuation per se did not appear to lead to supralaryngeal modifications. However, there was a clear distinction between the supralaryngeal articulation of words in broad focus and those in contrastive focus. We conclude that supralaryngeal articulation – in terms of acoustic duration and lip kinematics – is related directly to the expression of focus structure and contrastivity, and is not, contrary to conclusions drawn in previous studies, mediated by the presence or absence of accent.