Publications

Abstract

The ultimate goal of our research is to develop a computational model of human speech recognition that is able to capture the effects of fine-grained acoustic variation on speech recognition behaviour. As part of this work we are investigating automatic feature classifiers that are able to create reliable and accurate transcriptions of the articulatory behaviour encoded in the acoustic speech signal. In the experiments reported here, we compared support vector machines (SVMs) with multilayer perceptrons (MLPs). MLPs have been widely (and rather successfully) used for the task of multi-value articulatory feature classification, while (to the best of our knowledge) SVMs have not. This paper compares the performances of the two classifiers and analyses the results in order to better understand the articulatory representations. It was found that the MLPs outperformed the SVMs, but it is concluded that both classifiers exhibit similar behaviour in terms of patterns of errors.

Abstract

This paper addresses the issue of large-vocabulary recognition in a specific word class. We propose a two-pass strategy in which only major cities are explicitly represented in the first stage lexicon. An unknown word model encoded as a phone loop is used to detect OOV city names (referred to as rare city names). After which SpeM, a tool that can extract words and word-initial cohorts from phone graphs on the basis of a large fallback lexicon, provides an N-best list of promising city names on the basis of the phone sequences generated in the first stage. This N-best list is then inserted into the second stage lexicon for a subsequent recognition pass. Experiments were conducted on a set of spontaneous telephone-quality utterances each containing one rare city name. We tested the size of the N-best list and three types of language models (LMs). The experiments showed that SpeM was able to include nearly 85% of the correct city names into an N-best list of 3000 city names when a unigram LM, which also boosted the unigram scores of a city name in a given state, was used.

Abstract

A stretch of speech is often consistent with multiple words, e.g., the sequence /hæm/ is consistent with ‘ham’ but also with the first syllable of ‘hamster’, resulting in temporary ambiguity. However, to what degree does this lexical embedding occur? Analyses on two corpora of spoken Dutch showed that 11.9%-19.5% of polysyllabic word tokens have word-initial embedding, while 4.1%-7.5% of monosyllabic word tokens can appear word-initially embedded. This is much lower than suggested by an analysis of a large dictionary of Dutch. Speech processing thus appears to be simpler than one might expect on the basis of statistics on a dictionary.

Abstract

In this paper, we illustrate the close parallels between the research fields of human speech recognition (HSR) and automatic speech recognition (ASR) using a computational model of human word recognition, SpeM, which was built using techniques from ASR. We show that ASR has proven to be useful for improving models of HSR by relieving them of some of their shortcomings. However, in order to build an integrated computational model of all aspects of HSR, a lot of issues remain to be resolved. In this process, ASR algorithms and techniques definitely can play an important role.

Abstract

Evidence that listeners use durational cues to help resolve temporarily ambiguous speech input has accumulated over the past few years. In this paper, we investigate whether durational cues are also beneficial for word recognition in a computational model of spoken-word recognition. Two sets of simulations were carried out using the acoustic signal as input. The simulations showed that the computational model, like humans, takes benefit from durational cues during word recognition, and uses these to disambiguate the speech signal. These results thus provide support for the theory that durational cues play a role in spoken-word recognition.

Abstract

Good teachers know their students, and exploit this knowledge to adapt or optimise their instruction. Traditional teachers know their students because they interact with them face-to-face in classroom or one-to-one tutoring sessions. In these settings, they can build student models, i.e., by exploiting the multi-faceted nature of human-human communication. In distance-learning contexts, teacher and student have to cope with the lack of such direct interaction, and this must have detrimental effects for both teacher and student. In a past study we have analysed teacher requirements for tracking student actions in computer-mediated settings. Given the results of this study, we have devised and implemented a tool that allows teachers to keep track of their learners'interaction in e-learning systems. We present the tool's functionality and user interfaces, and an evaluation of its usability.

Abstract

The INTERSPEECH 2017 Computational Paralinguistics Challenge addresses three different problems for the first time in research competition under well-defined conditions: In the Addressee sub-challenge, it has to be determined whether speech produced by an adult is directed towards another adult or towards a child; in the Cold sub-challenge, speech under cold has to be told apart from ‘healthy’ speech; and in the Snoring subchallenge, four different types of snoring have to be classified. In this paper, we describe these sub-challenges, their conditions, and the baseline feature extraction and classifiers, which include data-learnt feature representations by end-to-end learning with convolutional and recurrent neural networks, and bag-of-audiowords for the first time in the challenge series

Abstract

This paper presents a study on the reduction of word-final [t]s in conversational standard Dutch. Based on a large amount of tokens annotated on the segmental level, we show that the bigram frequency and the segmental context are the main predictors for the absence of [t]s. In a second study, we present an analysis of the detailed acoustic properties of word-final [t]s and we show that bigram frequency and context also play a role on the subsegmental level. This paper extends research on the realization of /t/ in spontaneous speech and shows the importance of incorporating sub-segmental properties in models of speech.

Abstract

This presentation will address aspects of the expression of emotion in non-verbal vocal behaviour, specifically attempting to determine the roles of both positive and negative emotions, their acoustic bases, and the extent to which these are recognized in non-Western cultures.

Abstract

Speakers sometimes modify their gestures during the process of production into adaptors such as hair touching or eye scratching. Such disguised adaptors are evidence that the speaker can monitor their gestures. In this study, we investigated when and how disguised adaptors are first produced by children. Sixty elementary school children participated in this study (ten children in each age group; from 7 to 12 years old). They were instructed to watch a cartoon and retell it to their parents. The results showed that children did not produce disguised adaptors until the age of 8. The disguised adaptors accompany fluent speech until the children are 10 years old and accompany dysfluent speech until they reach 11 or 12 years of age. These results suggest that children start to monitor their gestures when they are 9 or 10 years old. Cognitive changes were considered as factors to influence emergence of disguised adaptors

Abstract

One of the many results of the "Space" research project conducted at the MPI for Psycholinguistics is that there are three "Frames of spatial Reference" (FoRs), the relative, the intrinsic and the absolute FoR. Cross-linguistic research showed that speakers who prefer one FoR in verbal spatial references rely on a comparable coding system for memorizing spatial configurations and for making inferences with respect to these spatial configurations in non-verbal problem solving. Moreover, research results also revealed that in some languages these verbal FoRs also influence gestural behavior. These results document the close interrelationship between language, culture and cognition in the domain "Space". The proper description of these interrelationships in the spatial domain requires language and culture specific ontologies.

Abstract

The present paper is part of a large research programme investigating the nature and properties of the predicate logic inherent in natural language. The general hypothesis is that natural speakers start off with a basic-natural logic, based on natural cognitive functions, including the basic-natural way of dealing with plural objects. As culture spreads, functional pressure leads to greater generalization and mathematical correctness, yielding ever more refined systems until the apogee of standard modern predicate logic. Four systems of predicate calculus are considered: Basic-Natural Predicate Calculus (BNPC), Aritsotelian-Abelardian Predicate Calculus (AAPC), Aritsotelian-Boethian Predicate Calculus (ABPC), also known as the classic Square of Opposition, and Standard Modern Predicate Calculus (SMPC). (ABPC is logically faulty owing to its Undue Existential Import (UEI), but that fault is repaired by the addition of a presuppositional component to the logic.) All four systems are checked against seven natural logical intuitions. It appears that BNPC scores best (five out of seven), followed by ABPC (three out of seven). AAPC and SMPC finish ex aequo with two out of seven.

Abstract

Turn-taking in conversation is a cognitively demanding process that proceeds rapidly due to interlocutors utilizing a range of cues
to aid prediction. In the present study we set out to test recent claims that content question words (also called wh-words) sound similar within languages as an adaptation to help listeners predict
that a question is about to be asked. We test whether upcoming questions can be predicted based on the first phoneme of a turn and the prior context. We analyze the Switchboard corpus of English
by means of a decision tree to test whether /w/ and /h/ are good statistical cues of upcoming questions in conversation. Based on the results, we perform a controlled experiment to test whether
people really use these cues to recognize questions. In both studies
we show that both the initial phoneme and the sequential context help predict questions. This contributes converging evidence that elements of languages adapt to pragmatic pressures applied during
conversation.

Abstract

One of the most pressing issues in dis-
continuous constituency transition-based
parsing is that the relevant information for
parsing decisions could be located in any
part of the stack or the buffer. In this pa-
per, we propose a solution to this prob-
lem by replacing the structured percep-
tron model with a recursive neural model
that computes a global representation of
the configuration, therefore allowing even
the most remote parts of the configura-
tion to influence the parsing decisions. We
also provide a detailed analysis of how
this representation should be built out of
sub-representations of its core elements
(words, trees and stack). Additionally, we
investigate how different types of swap or-
acles influence the results. Our model is
the first neural discontinuous constituency
parser, and it outperforms all the previ-
ously published models on three out of
four datasets while on the fourth it obtains
second place by a tiny difference.

Abstract

The problem of identifying and correcting confusibles, i.e. context-sensitive spelling errors, in text is typically tackled using specifically trained machine learning classifiers. For each different set of confusibles, a specific classifier is trained and tuned. In this research, we investigate a more generic approach to context-sensitive confusible correction. Instead of using specific classifiers, we use one generic classifier based on a language model. This measures the likelihood of sentences with different possible solutions of a confusible in place. The advantage of this approach is that all confusible sets are handled by a single model. Preliminary results show that the performance of the generic classifier approach is only slightly worse that that of the specific classifier approach

Abstract

In the context of confusible disambiguation (spelling correction that requires context), the synchronous back-off strategy combined with traditional n-gram language models performs well. However, when alternatives consist of a different number of tokens, this classification technique cannot be applied directly, because the computation of the probabilities is skewed. Previous work already showed that probabilities based on different order n-grams should not be compared directly. In this article, we propose new probability metrics in which the size of the n is varied according to the number of tokens of the confusible alternative. This requires access to n-grams of variable length. Results show that the synchronous back-off method is extremely robust. We discuss the use of suffix trees as a technique to store variable length n-gram information efficiently.

Abstract

Listeners can make perceptual adjustments to phoneme categories in response to a talker who consistently produces a specific phoneme ambiguously. We investigate here whether this type of perceptual learning is also used to adapt to regional accent differences. Listeners were exposed to words produced by a Flemish talker whose realization of [x℄or [h℄ was ambiguous (producing [x℄like [h℄is a property of the West-Flanders regional accent). Before and after exposure they categorized a [x℄-[h℄continuum. For both Dutch and Flemish listeners there was no shift of the categorization boundary after exposure to ambiguous sounds in [x℄- or [h℄-biasing contexts. The absence of a lexically-driven learning effect for this contrast may be because [h℄is strongly influenced by coarticulation. As is not stable across contexts, it may be futile to adapt its representation when new realizations are heard

Abstract

Motivated form-meaning mappings are pervasive in sign languages, and iconicity has recently been shown to facilitate sign learning from early on. This study investigated the role of iconicity for language acquisition in Turkish Sign Language (TID). Participants were 43 signing children (aged 10 to 45 months) of deaf parents. Sign production ability was recorded using the adapted version of MacArthur Bates Communicative Developmental Inventory (CDI) consisting of 500 items for TID. Iconicity and familiarity ratings for a subset of 104 signs were available. Our results revealed that the iconicity of a sign was positively correlated with the percentage of children producing a sign and that iconicity significantly predicted the percentage of children producing a sign, independent of familiarity or phonological complexity. Our results are consistent with previous findings on sign language acquisition and provide further support for the facilitating effect of iconic form-meaning mappings in sign learning.

Abstract

This paper describes ongoing research aiming at the description of variation in speech as represented by asynchronous articulatory features. We will first illustrate how distances in the articulatory feature space can be used for event detection along speech trajectories in this space. The temporal structure imposed by the cosine distance in articulatory feature space coincides to a large extent with the manual segmentation on phone level. The analysis also indicates that the articulatory feature representation provides better such alignments than the MFCC representation does. Secondly, we will present first results that indicate that articulatory features can be used to probe for acoustic differences in the onsets of Dutch singulars and plurals.

Abstract

This paper builds on previous work that aims at unraveling the structure of the speech signal by means of using probabilistic representations. The context of this work is a multi-pass speech recognition system in which a phone lattice is created and used as a basis for a lexical search in which symbolic mismatches are allowed at certain costs. The focus is on the optimization of the costs of phone insertions, deletions and substitutions that are used in the lexical decoding pass. Two optimization approaches are presented, one related to a multi-pass computational model for human speech recognition, the other based on a decoding in which Bayes’ risks are minimized. In the final section, the advantages of these optimization methods are discussed and compared.

Abstract

This paper investigates the interaction between acoustic scores and symbolic mismatch penalties in multi-pass speech decoding techniques that are based on the creation of a segment graph followed by a lexical search. The interaction between acoustic and symbolic mismatches determines to a large extent the structure of the search space of these multipass approaches. The background of this study is a recently developed computational model of human word recognition, called SpeM. SpeM is able to simulate human word recognition data and is built as a multi-pass speech decoder. Here, we focus on unravelling the structure of the search space that is used in SpeM and similar decoding strategies. Finally, we elaborate on the close relation between distances in this search space, and distance measures in search spaces that are based on a combination of acoustic and phonetic features.

Abstract

This paper investigates the processes in comprehending spoken noun-noun compounds, using data from the BALDEY database. BALDEY contains lexicality judgments and reaction times (RTs) for Dutch stimuli for which also linguistic information is included. Two different approaches are combined. The first is based on regression by Dynamic Survival Analysis, which models decisions and RTs as a consequence of the fact that a cumulative density function exceeds some threshold. The parameters of that function are estimated from the observed RT data. The second approach is based on DIANA, a process-oriented computational model of human word comprehension, which simulates the comprehension process with the acoustic stimulus as input. DIANA gives the identity and the number of the word candidates that are activated at each 10 ms time step.

Both approaches show how the processes involved in comprehending compounds change during a stimulus. Survival Analysis shows that the impact of word duration varies during the course of a stimulus. The density of word and non-word hypotheses in DIANA shows a corresponding pattern with different regimes. We show how the approaches complement each other, and discuss additional ways in which data and process models can be combined.

Abstract

The present study shows that [t] consonants are affected by probabilistic factors in a syllable-timed language as French, and in spontaneous as well as in journalistic speech. Study 1 showed a word bigram frequency effect in spontaneous French, but its exact nature depended on the corpus on which the probabilistic measures were based. Study 2 investigated journalistic speech and showed an effect of the joint frequency of the test word and its following word. We discuss the possibility that these probabilistic effects are due to the speaker’s planning of upcoming words, and to the speaker’s adaptation to the listener’s needs.

Abstract

Even advanced Spanish speakers of second language English tend to confuse the pronouns ‘he’ and ‘she’, often without even noticing their mistake (Lahoz, 1991). A study by AntónMéndez (2010) has indicated that a possible reason for this error is the fact that Spanish is a pro-drop language. In order to test this hypothesis, we used an extension of Dual-path (Chang, 2002), a computational cognitive model of sentence production, to simulate two models of bilingual speech production of second language English. One model had Spanish (ES) as a native language, whereas the other learned a Spanish-like language that used the pronoun at all times (non-pro-drop Spanish, NPD_ES). When tested on L2 English sentences, the bilingual pro-drop Spanish model produced significantly more gender pronoun errors, confirming that pronoun dropping could indeed be responsible for the gender confusion in natural language use as well.

Abstract

We investigated whether non-native listeners can exploit phonetic detail in recognizing potentially ambiguous utterances, as native listeners can [6, 7, 8, 9, 10]. Due to the phenomenon of intrusive /r/, the English phrase extra ice may sound like extra rice. A production study indicates that the intrusive /r/ can be distinguished from the onset /r/ in rice, as it is phonetically weaker. In two cross-modal identity priming studies, however, we found no conclusive evidence that Dutch learners of English are able to make use of this difference. Instead, auditory primes such as extra rice and extra ice with onset and intrusive /r/s activate both types of targets such as ice and rice. This supports the notion of spurious lexical activation in L2 perception.

Abstract

A dominant hypothesis in empirical research on the evolution of language is the following: the fundamental difference between animal and human communication systems is captured by the distinction between regular and more complex non-regular grammars. Studies reporting successful artificial grammar learning of nested recursive structures and imaging studies of the same have methodological shortcomings since they typically allow explicit problem solving strategies and this has been shown to account for the learning effect in subsequent behavioral studies. The present study overcomes these shortcomings by using subtle violations of agreement structure in a preference classification task. In contrast to the studies conducted so far, we use an implicit learning paradigm, allowing the time needed for both abstraction processes and consolidation to take place. Our results demonstrate robust implicit learning of recursively embedded structures (context-free grammar) and recursive structures with cross-dependencies (context-sensitive grammar) in an artificial grammar learning task spanning 9 days. Keywords: Implicit artificial grammar learning; centre embedded; cross-dependency; implicit learning; context-sensitive grammar; context-free grammar; regular grammar; non-regular grammar

Abstract

This paper describes a new flexible delexicalization method based on glottal excited parametric speech synthesis scheme. The system utilizes inverse filtered glottal flow and all-pole modelling of the vocal tract. The method provides a possibility to retain and manipulate all relevant prosodic features of any kind of speech. Most importantly, the features include voice quality, which has not been properly modeled in earlier delexicalization methods. The functionality of the new method was tested in a prosodic tagging experiment aimed at providing word prominence data for a text-to-speech synthesis system. The experiment confirmed the usefulness of the method and further corroborated earlier evidence that linguistic factors influence the perception of prosodic prominence.

Abstract

The present study compared the role of metrical stress in comprehension and production of three-year-old children with a familial risk of dyslexia with that of normally developing children. A visual fixation task with stress (mis-)matches in bisyllabic words, as well as a non-word repetition task with bisyllabic targets were presented to the control and at-risk children. Results show that the at-risk group is less sensitive to stress mismatches in word recognition than the control group. Correct production of metrical stress patterns did not differ significantly between the groups, but the percentages of phonemes produced correctly were lower for the at-risk than the control group. The findings indicate that processing of metrical stress patterns is not impaired in at-risk children, but that the at-risk group cannot exploit metrical stress in word recognition

Abstract

This corpus study investigates how children figure out that functional modals
like must can express various flavors of modality. We examine how modality is
expressed in speech to and by children, and find that the way speakers use
modals may obscure their polysemy. Yet, children eventually figure it out. Our
results suggest that some do before age 3. We show that while root and
epistemic flavors are not equally well-represented in the input, there are robust
correlations between flavor and aspect, which learners could exploit to discover
modal polysemy.

Abstract

A not uncommon assumption in psycholinguistics is that the N400 directly indexes high-level semantic integration, the compositional, word-driven construction of sentence- and discourse-level meaning in some language-relevant unification space. The various discourse- and speaker-dependent modulations of the N400 uncovered by us and others are often taken to support this 'compositional integration' position. In my talk, I will argue that these N400 modulations are probably better interpreted as only indirectly reflecting compositional sense-making. The account that I will advance for these N400 effects is a variant of the classic Kutas and Federmeier (2002, TICS) memory retrieval account in which context effects on the word-elicited N400 are taken to reflect contextual priming of LTM access. It differs from the latter in making more explicit that the contextual cues that prime access to a word's meaning in LTM can range from very simple (e.g., a single concept) to very complex ones (e.g., a structured representation of the current discourse). Furthermore, it incorporates the possibility, suggested by recent N400 findings, that semantic retrieval can also be intensified in response to certain ‘relevance signals’, such as strong value-relevance, or a marked delivery (linguistic focus, uncommon choice of words, etc). In all, the perspective I'll draw is that in the context of discourse-level language processing, N400 effects reflect an 'overlay of technologies', with the construction of discourse-level representations riding on top of more ancient sense-making technology.

Abstract

Listeners require context to understand the casual pronunciation variants of words that are typical of spontaneous speech (Ernestus et al., 2002). The present study reports two auditory lexical decision experiments, investigating listeners' use of semantic contextual information in the comprehension of unreduced and reduced words. We found a strong semantic priming effect for low frequency unreduced words, whereas there was no such effect for reduced words. Word frequency was facilitatory for all words. These results show that semantic context is relevant especially for the comprehension of unreduced words, which is unexpected given the listener driven explanation of reduction in spontaneous speech.

Abstract

Lexical recognition is typically slower in L2 than in L1. Part of the difficulty comes from a not precise enough processing of L2 phonemes. Consequently, L2 listeners fail to eliminate candidate words that L1 listeners can exclude from competing for recognition. For instance, the inability to distinguish /r/ from /l/ in rocket and locker makes for Japanese listeners both words possible candidates when hearing their onset (e.g., Cutler, Weber, and Otake, 2006). The L2 disadvantage can, however, be dispelled: For L2 listeners, but not L1 listeners, L2 speech from a non-native talker with the same language background is known to be as intelligible as L2 speech from a native talker (e.g., Bent and Bradlow, 2003). A reason for this may be that L2 listeners have ample experience with segmental deviations that are characteristic for their own accent. On this account, only phonemic deviations that are typical for the listeners’ own accent will cause spurious lexical activation in L2 listening (e.g., English magic pronounced as megic for Dutch listeners). In this talk, I will present evidence from cross-modal priming studies with a variety of L2 listener groups, showing how the processing of phonemic deviations is accent-specific but withstands fine phonetic differences.

Abstract

In a visual-world study, we examined spoken-wordrecognition in Spanish. Spanish listeners followed spoken instructions to click on pictures while their eye movements were monitored. When instructed to click on the picture of a door (puerta), they experienced interference from the picture of a pig (p u e r c o ). The same interference from phonologically related items was observed when the displays contained printed names or a combination of pictures with their names printed underneath, although the effect was strongest for displays with printed names. Implications of the finding that the interference effect can be induced with standard pictorial displays as well as with orthographic displays are discussed.

Abstract

Utilization of computer tools in linguistic research has gained importance with the maturation of media frameworks for the handling of digital audio and video. The increased use of these tools in gesture, sign language and multimodal interaction studies has led to stronger requirements on the flexibility, the efficiency and in particular the time accuracy of annotation tools. This paper describes the efforts made to make ELAN a tool that meets these requirements, with special attention to the developments in the area of time accuracy. In subsequent sections an overview will be given of other enhancements in the latest versions of ELAN, that make it a useful tool in multimodality research.

Abstract

A number of serious reasons will convince an increasing amount of researchers to store their relevant material in centers which we will call "language resource archives". They combine the duty of taking care of long-term preservation as well as the task to give access to their material to different user groups. Access here is meant in the sense that an active interaction with the data will be made possible to support the integration of new data, new versions or commentaries of all sort. Modern Language Resource Archives will have to adhere to a number of basic principles to fulfill all requirements and they will have to be involved in federations to create joint language resource domains making it even more simple for the researchers to access the data. This paper makes an attempt to formulate the essential pillars language resource archives have to adhere to.

Abstract

Set is widely used as a kind of basic data structure. However, when it is used for large scale data set the cost of storage, search and transport is overhead. The bloom filter uses a fixed size bit string to represent elements in a static set, which can reduce storage space and search cost that is a fixed constant. The time-space efficiency is achieved at the cost of a small probability of false positive in membership query. However, for many applications the space savings and locating time constantly outweigh this drawback. Dynamic bloom filter (DBF) can support concisely representation and approximate membership queries of dynamic set instead of static set. It has been proved that DBF not only possess the advantage of standard bloom filter, but also has better features when dealing with dynamic set. This paper proposes a time-dependent multiple bit-strings bloom filter (TMBF) which roots in the DBF and targets on dynamic incremental set. TMBF uses multiple bit-strings in time order to present a dynamic increasing set and uses backward searching to test whether an element is in a set. Based on the system logs from a real P2P file sharing system, the evaluation shows a 20% reduction in searching cost compared to DBF.