Paul Boersma

Empirical Tests of the Gradual Learning Algorithm

The Gradual Learning Algorithm (Boersma 1997) is a constraint-ranking algorithm for learning optimality-theoretic grammars. The purpose of this article is to assess the capabilities of the Gradual Learning Algorithm, particularly in comparison with the Constraint Demotion algorithm of Tesar and Smolensky (1993, 1996, 1998, 2000), which initiated the learnability research program for Optimality Theory. We argue that the Gradual Learning Algorithm has a number of special advantages: it can learn free variation, deal effectively with noisy learning data, and account for gradient well-formedness judgments. The case studies we examine involve Ilokano reduplication and metathesis, Finnish genitive plurals, and the distribution of English light and dark /l/.

Bridging the Gap between L2 Speech Perception Research and Phonological Theory.

A series of experiments shows that Spanish learners of English acquire the ship-sheep contrast in a way specific to their target dialect (Scottish or Southern British English) and that many learners exhibit a perceptual strategy found in neither Spanish nor English. To account for these facts as well as for the findings of earlier research on second language (L2) speech perception, we provide an Optimality Theoretic model of phonological categorization that comes with a formal learning algorithm for its acquisition. Within this model, the dialect-dependent and L2-specific facts provide evidence for the hypotheses of Full Transfer and Full Access. We would like to thank the audiences of the 25th Penn Linguistics Colloquium (Philadelphia, 2001), Generative Approaches to Language Acquisition 2001 (Palmela), and EuroSLA 11 (Paderborn, 2001) and of talks in Lisbon, Amsterdam, Edinburgh, Utrecht, and Munich, all in 2001 as well, for their questions and remarks, and Rachel Hayes, Michael Sharwood Smith, and five anonymous SSLA reviewers for their helpful comments on earlier versions of this paper. This research was partially sponsored by grant 355-75-003 to Boersma from the Netherlands Organization for Scientific Research (NWO).

A cross-dialect acoustic description of vowels: Brazilian and European Portuguese

This paper examines four acoustic correlates of vowel identity in Brazilian Portuguese (BP) and European Portuguese (EP): first formant (F1), second formant (F2), duration, and fundamental frequency (F0). Both varieties of Portuguese display some cross-linguistically common phenomena: vowel-intrinsic duration, vowel-intrinsic pitch, gender-dependent size of the vowel space, gender-dependent duration, and a skewed symmetry in F1 between front and back vowels. Also, the average difference between the vocal tract sizes associated with /i/ and /u/, as measured from formant analyses, is comparable to the average difference between male and female vocal tract sizes. A language-specific phenomenon is that in both varieties of Portuguese the vowel-intrinsic duration effect is larger than in many other languages. Differences between BP and EP are found in duration (BP has longer stressed vowels than EP), in F1 (the lower-mid front vowel approaches its higher-mid counterpart more closely in EP than in BP), and in the size of the intrinsic pitch effect (larger for BP than for EP).

Context-specific acoustic differences between Peruvian and Iberian Spanish vowels

This paper examines four acoustic properties (duration F0, F1, and F2) of the monophthongal vowels of Iberian Spanish (IS) from Madrid and Peruvian Spanish (PS) from Lima in various consonantal contexts (/s/, /f/, /t/, /p/, and /k/) and in various phrasal contexts (in isolated words and sentence-internally). Acoustic measurements on 39 speakers, balanced by dialect and gender, can be generalized to the following differences between the two dialects. The vowel /a/ has a lower first formant in PS than in IS by 6.3%. The vowels /e/ and /o/ have more peripheral second-formant (F2) values in PS than in IS by about 4%. The consonant /s/ causes more centralization of the F2 of neighboring vowels in IS than in PS. No dialectal differences are found for the effect of phrasal context. Next to the between-dialect differences in the vowels, the present study finds that /s/ has a higher spectral center of gravity in PS than in IS by about 10%, that PS speakers speak slower than IS speakers by about 9%, and that Spanish-speaking women speak slower than Spanish-speaking men by about 5% (irrespective of dialect).

Spectral characteristics of three styles of Croatian folk singing

This paper examines the differences between three Croatian folk singing styles, namely klapa, ojkanje, and tarankanje. In order to factor out singer-specific properties, each of the styles was performed by the same 12 professional male singers. The 36 performances were analyzed with a long-term average spectrum (LTAS) method from which direct effects of the pitch distribution were removed. After factoring out each singers average, the 36 pitch-corrected LTAS contours were reduced to a two-dimensional representation in two ways: (1) a principal-component analysis and (2) a graphical plot of spectral slope versus speakers formant strength. Both ways clearly separate the three styles. The spectrum of the klapa style turns out to be similar to that of speech. The ojkanje style is extremely loud and shows two spectral peaks: a sharp one tuned at twice the fundamental frequency and appropriate for long-distance communication on mountain slopes, and a broad one around 3.5 kHz, reminiscent of a speakers formant. The tarankanje style has a very flat spectrum implemented by vocal pressedness and nasality, which is appropriate for blending into or imitating the timbral characteristics of the sopile folk instrument.

Should jitter be measured by peak picking or by waveform matching

Disclaimer/Complaints regulations If you believe that digital publication of certain material infringes any of your rights or (privacy) interests, please let the Library know, stating your reasons. In case of a legitimate complaint, the Library will make the material inaccessible and/or remove it from the website. Please Ask the Library: https://uba.uva.nl/en/contact, or a letter to: Library of the University of Amsterdam, Secretariat, Singel 425, 1012 WP Amsterdam, The Netherlands. You will be contacted as soon as possible.

Some Correct Error-Driven Versions of the Constraint Demotion Algorithm

This article shows that Error-Driven Constraint Demotion (EDCD), an error-driven learning algorithm proposed by Tesar (1995) for Prince and Smolenskys (1993/2004) version of Optimality Theory, can fail to converge to a correct totally ranked hierarchy of constraints, unlike the earlier non-error-driven learning algorithms proposed by Tesar and Smolensky (1993). The cause of the problem is found in Tesars use of mark-pooling ties, indicating that EDCD can be repaired by assuming Anttilas (1997) permuting ties instead. Proofs show, and simulations confirm, that totally ranked hierarchies can indeed be found by both this repaired version of EDCD and Boersmas (1998) Minimal Gradual Learning Algorithm.

Fast phonetic learning occurs already in 2-to-3-month old infants. An ERP study

An important mechanism for learning speech sounds in the first year of life is “distributional learning,” i.e., learning by simply listening to the frequency distributions of the speech sounds in the environment. In the lab, fast distributional learning has been reported for infants in the second half of the first year; the present study examined whether it can also be demonstrated at a much younger age, long before the onset of language-specific speech perception (which roughly emerges between 6 and 12 months). To investigate this, Dutch infants aged 2 to 3 months were presented with either a unimodal or a bimodal vowel distribution based on the English /æ/~/ε/ contrast, for only 12 minutes. Subsequently, mismatch responses (MMRs) were measured in an oddball paradigm, where one half of the infants in each group heard a representative [æ] as the standard and a representative [ε] as the deviant, and the other half heard the same reversed. The results (from the combined MMRs during wakefulness and active sleep) disclosed a larger MMR, implying better discrimination of [æ] and [ε], for bimodally than unimodally trained infants, thus extending an effect of distributional training found in previous behavioral research to a much younger age when speech perception is still universal rather than language-specific, and to a new method (using event-related potentials). Moreover, the analysis revealed a robust interaction between the distribution (unimodal vs. bimodal) and the identity of the standard stimulus ([æ] vs. [ε]), which provides evidence for an interplay between a perceptual asymmetry and distributional learning. The outcomes show that distributional learning can affect vowel perception already in the first months of life.

Spreading in functional phonology

The occurrence of and the restrictions on the temporal spreading of phonological feature values (assimilation, harmony) are the results of interactions between the functional principles of minimizing articulatory effort and minimizing perceptual confusion. This proposal is tested on the typology of opacity to nasal spreading. While the sonority approach of Gnanadesikan (1995) meets with insuperable problems with regard to the position of /h/ in the hierarchy, and the feature-geometric representational approach of Piggott (1992) needs to take recourse to ad-hoc conditions in UG in order to get the hierarchy right, the functional approach accurately predicts the attested typology.

The odds of eternal optimization in optimality theory

The first part of this paper shows that a non-teleological account of sound change is possible if we assume two things: first, that Optimality-Theoretic constraints that do not contribute to determining the winning candidate are ranked randomly with respect to each other, i.e. differently for every speaker; second, that learners acquire as their underlying representations the forms that they detect most often in their environment. The resulting variation-and-selection scheme can be regarded as locally optimizing. It is shown, however, that it is possible that a sequence of such optimizing sound changes ends up in a loop rather than in a single absorbing final state. This kind of cyclic optimization is shown to be exactly what happened in the attested and reconstructed changes in the Indo-European consonant systems. The second part of this paper presents a simulation that shows that cyclic optimization is not only possible but also rather likely: twenty percent of all inventories are in an optimizing loop or heading towards one.