Abstract not found

ally - positing boundaries inside phoneme transitions which are unlikely to occur word-internally. This has implications for complex words. If the phonology across a morpheme boundary is unlikely to occur morpheme-internally, then the preprocessor will posit a boundary, and so facilitate decomposition. For example the /pf/ transition in pipeful is unlikely to occur within a simple word in English. The whole-word route will therefore be disadvantaged relative to the decomposed route, because it doesn't align with hypothesized boundaries. Three experiments demonstrate that phonotactics are relevant to morphological decomposition. In experiment 1, a simple recurrent network was trained to spot the boundaries between monomorphemes, and then tested on multimorphemic words. Despite never having previously encountered a morphologically complex word, the network hypothesized a boundary at 60% of word-internal morpheme boundaries. This demonstrates that English is conf

This paper explores the perception and well-formedness of nonsense words containing nasal-obstruent (NO) clusters. Morpheme internally, these clusters are subject to a homorganicity constraint in English, which would be represented in a conventional phonology by a feature spreading rule. Yet such a rule does not do justice to the lexical statistics. The strength of the homorganicity requirement depends on the manner of the obstruent and the place of articulation of both the nasal and the obstruent. Some NO clusters are therefore extremely frequent (e.g. /nt/), others are unattested (/m7/) , and yet others fall between these two extremes (/nf/). Because NO clusters are a phonetically coherent set, and sample the range of frequencies finely, they make Hay, Pierrehumbert and Beckman

Abstract not found

A series of three experiments investigated semantic and syntactic effects in the production of Adjective+Noun phrases in Dutch. Bilinguals (Dutch native speakers) were presented with English nouns and were asked to produce an Adjective+Noun phrase in Dutch which included the translation of the noun. In two experiments, we blocked items by either semantic category or grammatical gender.We found that participants performed the task slower when the target nouns were of the same semantic category than when they were from different categories; and faster when they were of the same grammatical gender than when they were of different gender. In a final experiment, both manipulations were crossed in order to both replicate the previous experiments and to test for interactions between the two effects. The results of the first two experiments were replicated, and crucially no interaction was found. These findings are compatible with models of lexical retrieval in production in which, first lexico-semantic and lexico-syntactic information are separable; second the flow of activation between the two is feedforward.

Abstract not found

Four experiments investigated potential influences of spelling on single word speech production. A form-preparation paradigm that showed priming effects for words with initial form overlap was used to investigate whether words with form overlap, but different spelling (e.g., "camel"-"kidney") also show priming. Experiment 1 demonstrated that such words did not benefit from the form overlap, suggesting that the incongruent spelling disrupted the form-preparation effect. Experiment 2 replicated the first experiment with an independent set of items and an improved design, and once again showed a disruptive effect of spelling. To divert participants' attention from the spelling of the targets, Experiment 3 was conducted entirely in the auditory domain, but yielded the same outcome as before. Experiment 4 showed that matching initial letters alone, in the absence of matching sounds (e.g., "cycle"-"cobra"), did not produce priming. These findings raise the possibility that orthographic codes are mandatorily activated in speech production by literate speakers.

this paper is broader than standard, encompassing phoneme oppositions (binary or not), distinctive features (again, binary or not), suprasegmental features and even phonological rules such as phoneme deletion in certain contexts. This permits researchers with the appropriate corpora to answer questions like these: Is it more important to correctly hear the tone or the vowel in Cantonese? 8 Does Hindi make more use of aspiration or voicing? How much information is lost due to vowel reduction in unstressed syllables? If second-language speakers have trouble learning contrasts that are not present in their native language, e.g. the [l]-[r] distinction in English for Japanese speakers, how badly o are they?

The main goal of the speech synthesis group in SmartKom was to develop a natural sounding synthetic voice for the avatar “Smartakus ” that is judged to be agreeable, intelligible, and friendly by the users of the SmartKom system. Two aspects of the SmartKom scenario facilitate the achievement of this goal. First, since speech output is mainly intended for the interaction of Smartakus with the user, most of the output corresponds to dialog turns generated by the language generation module (see Chapter??). Therefore, most speech output can be generated from linguistic concepts produced by the language generation module (“concept-to-speech synthesis”, CTS) instead of from raw text (“text-to-speech synthesis”, TTS). The advantage of CTS over TTS is that it avoids errors that may be introduced by linguistic analysis in TTS mode. Second, the CTS approach narrows down the SmartKom synthesis domain from a theoretically open domain to a restricted domain, which makes unit selection synthesis a promising alternative to diphone synthesis for the SmartKom application. Multimodality introduces additional requirements for the synthesis module. The

Unsupervised learning algorithms based on Expectation Maximization (EM) are often straightforward to implement and provably converge on a local likelihood maximum. However, these algorithms often do not perform well in practice. Common wisdom holds that they yield poor results because they are overly sensitive to initial parameter values and easily get stuck in local (but not global) maxima. We present a series of experiments indicating that for the task of learning syllable structure, the initial parameter weights are not crucial. Rather, it is the choice of model class itself that makes the difference between successful and unsuccessful learning. We use a language-universal rule-based algorithm to find a good set of parameters, and then train the parameter weights using EM. We achieve word accuracy of 95.9 % on German and 97.1 % on English, as compared to 97.4 % and 98.1% respectively for supervised training. 1