Publications

Abstract

Word count estimation (WCE) from audio recordings has a number of applications, including quantifying the amount of speech that language-learning infants hear in their natural environments, as captured by daylong recordings made with devices worn by infants. To be applicable in a wide range of scenarios and also low-resource domains, WCE tools should be extremely robust against varying signal conditions and require minimal access to labeled training data in the target domain. For this purpose, earlier work has used automatic syllabification of speech, followed by a least-squares-mapping of syllables to word counts. This paper compares a number of previously proposed syllabifiers in the WCE task, including a supervised bi-directional long short-term memory (BLSTM) network that is trained on a language for which high quality syllable annotations are available (a “high resource language”), and reports how the alternative methods compare on different languages and signal conditions. We also explore additive noise and varying-channel data augmentation strategies for BLSTM training, and show how they improve performance in both matching and mismatching languages. Intriguingly, we also find that even though the BLSTM works on languages beyond its training data, the unsupervised algorithms can still outperform it in challenging signal conditions on novel languages.

Abstract

For almost two decades, the poor performance observed with the so-called Director task has been interpreted as evidence of limited use of Theory of Mind in communication. Here we propose a probabilistic model of common ground in referential communication that derives three inferences from an utterance: what the speaker is talking about in a visual context, what she knows about the context, and what referential expressions she prefers. We tested our model by comparing its inferences with those made by human participants and found that it closely mirrors their judgments, whereas an alternative model compromising the hearer’s expectations of cooperativeness and efficiency reveals a worse fit to the human data. Rather than assuming that common ground is fixed in a given exchange and may or may not constrain reference resolution, we show how common ground can be inferred as part of the process of reference assignment.

Abstract

While there are many studies on information retrieval models using full-text, there are presently no comparison studies of full-text retrieval vs. retrieval only over the titles of documents. On the one hand, the full-text of documents like scientific papers is not always available due to, e.g., copyright policies of academic publishers. On the other hand, conducting a search based on titles alone has strong limitations. Titles are short and therefore may not contain enough information to yield satisfactory search results. In this paper, we compare different retrieval models regarding their search performance on the full-text vs. only titles of documents. We use different datasets, including the three digital library datasets: EconBiz, IREON, and PubMed. The results show that it is possible to build effective title-based retrieval models that provide competitive results comparable to full-text retrieval. The difference between the average evaluation results of the best title-based retrieval models is only 3% less than those of the best full-text-based retrieval models.

Abstract

A series of experiment tested Ekman’s (1992) hypothesis that there are a set of positive basic emotions that are expressed using vocal para-linguistic sounds, e.g. laughter and cheers. The proposed categories investigated were amusement, contentment, pleasure, relief and triumph. Behavioural testing using a forced-choice task indicated that participants were able to reliably recognize vocal expressions of the proposed emotions. A cross-cultural study in the preliterate Himba culture in Namibia confirmed that these categories are also recognized across cultures. A recognition test of acoustically manipulated emotional vocalizations established that the recognition of different emotions utilizes different vocal cues, and that these in turn differ from the cues used when comprehending speech. In a study using fMRI we found that relative to a signal correlated noise baseline, the paralinguistic expressions of emotion activated bilateral superior temporal gyri and sulci, lateral and anterior to primary auditory cortex, which is consistent with the processing of non linguistic vocal cues in the auditory ‘what’ pathway. Notably amusement was associated with greater activation extending into both temporal poles and amygdale and insular cortex. Overall, these results support the claim that ‘happiness’ can be fractionated into amusement, pleasure, relief and triumph.

Abstract

Fine-Tracker is a speech-based model of human speech
recognition. While previous work has shown that Fine-Tracker
is successful at modelling aspects of human spoken-word
recognition, its speech recognition performance is not
comparable to that of human performance, possibly due to
suboptimal intermediate articulatory feature (AF)
representations. This study investigates the effect of improved
AF representations, obtained using a state-of-the-art deep
convolutional network, on Fine-Tracker’s simulation and
recognition performance: Although the improved AF quality
resulted in improved speech recognition; it, surprisingly, did
not lead to an improvement in Fine-Tracker’s simulation power.

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

We have recently developed a new model of human speech recognition, based on automatic speech recognition techniques [1]. The present paper has two goals. First, we show that the new model performs well in the recognition of lexically ambiguous input. These demonstrations suggest that the model is able to operate in the same optimal way as human listeners. Second, we discuss how to relate the behaviour of a recogniser, designed to discover the optimum path through a word lattice, to data from human listening experiments. We argue that this requires a metric that combines both path-based and word-based measures of recognition performance. The combined metric varies continuously as the input speech signal unfolds over time.

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

In this paper, we present a novel computational model of human speech recognition – called SpeM – based on the theory underlying Shortlist. We will show that SpeM, in combination with an automatic phone recogniser (APR), is able to simulate the human speech recognition process from the acoustic signal to the ultimate recognition of words. This joint model takes an acoustic speech file as input and calculates the activation flows of candidate words on the basis of the degree of fit of the candidate words with the input. Experiments showed that SpeM outperforms Shortlist on the recognition of ‘real-life’ input. Furthermore, SpeM performs only slightly worse than an off-the-shelf full-blown automatic speech recogniser in which all words are equally probable, while it provides a transparent computationally elegant paradigm for modelling word activations in human word recognition.

Abstract

This study investigated the encoding of metrical information during speech production in Dutch. In Experiment 1, participants were asked to judge whether bisyllabic picture names had initial or final stress. Results showed significantly faster decision times for initially stressed targets (e.g., LEpel 'spoon') than for targets with final stress (e.g., liBEL 'dragon fly'; capital letters indicate stressed syllables) and revealed that the monitoring latencies are not a function of the picture naming or object recognition latencies to the same pictures. Experiments 2 and 3 replicated the outcome of the first experiment with bi- and trisyllabic picture names. These results demonstrate that metrical information of words is encoded rightward incrementally during phonological encoding in speech production. The results of these experiments are in line with Levelt's model of phonological encoding.

Abstract

Three experiments examined whether infants recognise functors in phrases, and whether their representations of functors are phonetically well specified. Eight- and 13- month-old English infants heard monosyllabic lexical words preceded by real functors (e.g., the, his) versus nonsense functors (e.g., kuh); the latter were minimally modified segmentally (but not prosodically) from real functors. Lexical words were constant across conditions; thus recognition of functors would appear as longer listening time to sequences with real functors. Eightmonth- olds' listening times to sequences with real versus nonsense functors did not significantly differ, suggesting that they did not recognise real functors, or functor representations lacked phonetic specification. However, 13-month-olds listened significantly longer to sequences with real functors. Thus, somewhere between 8 and 13 months of age infants learn familiar functors and represent them with segmental detail. We propose that accumulated frequency of functors in input in general passes a critical threshold during this time.

Abstract

We report an individual with music-odor synaesthesia who experiences automatic and vivid odor sensations when she hears music. S’s odor associations were recorded on two days, and compared with those of two control participants. Overall, S produced longer descriptions, and her associations were of multiple odors at once, in comparison to controls who typically reported a single odor. Although odor associations were qualitatively different between S and controls, ratings of the consistency of their descriptions did not differ. This demonstrates that crossmodal associations between music and odor exist in non-synaesthetes too. We also found that S is better at discriminating between odors than control participants, and is more likely to experience emotion, memories and evaluations triggered by odors, demonstrating the broader impact of her synaesthesia.

Abstract

Sequences of reaction times (RT) produced by participants in an experiment are not only influenced by the stimuli, but by many other factors as well, including fatigue, attention, experience, IQ, handedness, etc. These confounding factors result in longterm effects (such as a participant’s overall reaction capability) and in short- and medium-time fluctuations in RTs (often referred to as ‘local speed effects’). Because stimuli are usually presented in a random sequence different for each participant, local speed effects affect the underlying ‘true’ RTs of specific trials in different ways across participants. To be able to focus statistical analysis on the effects of the cognitive process under study, it is necessary to reduce the effect of confounding factors as much as possible. In this paper we propose and compare techniques and criteria for doing so, with focus on reducing (‘filtering’) the local speed effects. We show that filtering matters substantially for the significance analyses of predictors in linear mixed effect regression models. The performance of filtering is assessed by the average between-participant correlation between filtered RT sequences and by Akaike’s Information Criterion, an important measure of the goodness-of-fit of linear mixed effect regression models.

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

The recent advent of deep learning techniques in speech tech-nology and in particular in automatic speech recognition hasyielded substantial performance improvements. This suggeststhat deep neural networks (DNNs) are able to capture structurein speech data that older methods for acoustic modeling, suchas Gaussian Mixture Models and shallow neural networks failto uncover. In image recognition it is possible to link repre-sentations on the first couple of layers in DNNs to structuralproperties of images, and to representations on early layers inthe visual cortex. This raises the question whether it is possi-ble to accomplish a similar feat with representations on DNNlayers when processing speech input. In this paper we presentthree different experiments in which we attempt to untanglehow DNNs encode speech signals, and to relate these repre-sentations to phonetic knowledge, with the aim to advance con-ventional phonetic concepts and to choose the topology of aDNNs more efficiently. Two experiments investigate represen-tations formed by auto-encoders. A third experiment investi-gates representations on convolutional layers that treat speechspectrograms as if they were images. The results lay the basisfor future experiments with recursive networks.

Abstract

We estimate lexical Concreteness for millions of words across 77 languages. Using a simple regression framework, we combine vector-based models of lexical semantics with experimental norms of Concreteness in English and Dutch. By applying techniques to align vector-based semantics across distinct languages, we compute and release Concreteness estimates at scale in numerous languages for which experimental norms are not currently available. This paper lays out the technique and its efficacy. Although this is a difficult dataset to evaluate immediately, Concreteness estimates computed from English correlate with Dutch experimental norms at $\rho$ = .75 in the vocabulary at large, increasing to $\rho$ = .8 among Nouns. Our predictions also recapitulate attested relationships with word frequency. The approach we describe can be readily applied to numerous lexical measures beyond Concreteness

Abstract

Do all languages convey semantic knowledge in the same way? If language simply mirrors the structure of the world, the answer should be a qualified “yes”. If, however, languages impose structure as much as reflecting it, then even ostensibly the “same” word in different languages may mean quite different things. We provide a first pass at a large-scale quantification of cross-linguistic semantic alignment of approximately 1000 meanings in 55 languages. We find that the translation equivalents in some domains (e.g., Time, Quantity, and Kinship) exhibit high alignment across languages while the structure of other domains (e.g., Politics, Food, Emotions, and Animals) exhibits substantial cross-linguistic variability. Our measure of semantic alignment correlates with known phylogenetic distances between languages: more phylogenetically distant languages have less semantic alignment. We also find semantic alignment to correlate with cultural distances between societies speaking the languages, suggesting a rich co-adaptation of language and culture even in domains of experience that appear most constrained by the natural world

Abstract

In situated communication, reference to an entity in the shared visual context can be established using eitheranexpression that conveys precise (minimally specified) or redundant (over-specified) information. There is, however, along-lasting debate in psycholinguistics concerningwhether the latter hinders referential processing. We present evidence from an eyetrackingexperiment recordingfixations as well asthe Index of Cognitive Activity –a novel measure of cognitive workload –supporting the view that over-specifications facilitate processing. We further present originalevidence that, above and beyond the effect of specificity,referring expressions thatuniformly reduce referential entropyalso benefitprocessing

Abstract

The task of automatic playlist continuation is generating a list of recommended tracks that can be added to an existing playlist. By suggesting appropriate tracks, i. e., songs to add to a playlist, a recommender system can increase the user engagement by making playlist creation easier, as well as extending listening beyond the end of current playlist. The ACM Recommender Systems Challenge 2018 focuses on such task. Spotify released a dataset of playlists, which includes a large number of playlists and associated track listings. Given a set of playlists from which a number of tracks have been withheld, the goal is predicting the missing tracks in those playlists. We participated in the challenge as the team Unconscious Bias and, in this paper, we present our approach. We extend adversarial autoencoders to the problem of automatic playlist continuation. We show how multiple input modalities, such as the playlist titles as well as track titles, artists and albums, can be incorporated in the playlist continuation task.

Abstract

This paper reports two experiments with vowels and consonants as phoneme detection targets in real words. In the first experiment, two relatively distinct vowels were compared with two confusible stop consonants. Response times to the vowels were longer than to the consonants. Response times correlated negatively with target phoneme length. In the second, two relatively distinct vowels were compared with their corresponding semivowels. This time, the vowels were detected faster than the semivowels. We conclude that response time differences between vowels and stop consonants in this task may reflect differences between phoneme categories in the variability of tokens, both in the acoustic realisation of targets and in the' representation of targets by subjects.

Abstract

Pitch, intonation, and speech rate were analyzed in a collection of everyday speech heard by one Dutch infant between the ages of six and nine months. Components of each of these variables were measured in the speech of three adult speakers (mother, father, baby-sitter) when they addressed the infant, and when they addressed another adult. The results are in line with previously reported findings which are usually based on laboratory or prearranged settings: infant-directed speech in a natural setting exhibits more pitch variation, a larger number of simple intonation contours, and slower speech rate than does adult-directed speech.

Abstract

Dutch distinguishes at least four sentence types: statements and questions, the latter type being subdivided into wh-questions (beginning with a question word), yes/no-questions (with inversion of subject and finite), and declarative questions (lexico-syntactically identical to statement). Acoustically, each of these (sub)types was found to have clearly distinct global F0-patterns, as well as a characteristic distribution of final rises [1,2]. The present paper explores the separate contribution of parameters of global downtrend and size of accent-lending pitch movements versus aspects of the terminal rise to the human identification of the four sentence (sub)types, at various positions in the time-course of the utterance. The results show that interrogativity in Dutch can be identified at an early point in the utterance. However, wh-questions are not distinct from statements.

Abstract

Before infants become mature speakers of their native language, they must acquire a robust word-recognition system which allows them to strike the balance between allowing some variation (mood, voice, accent) and recognizing variability that potentially changes meaning (e.g. cat vs hat). The current meta-analysis quantifies how the latter, termed mispronunciation sensitivity, changes over infants’ first three years, testing competing predictions of mainstream language acquisition theories. Our results show that infants were sensitive to mispronunciations, but accepted them as labels for target objects. Interestingly, and in contrast to predictions of mainstream theories, mispronunciation sensitivity was not modulated by infant age, suggesting that a sufficiently flexible understanding of native language phonology is in place at a young age.

Abstract

This study investigated the perception of American English phonemes by native listeners. Listeners identified either the consonant or the vowel in all possible English CV and VC syllables. The syllables were embedded in multispeaker babble at three signal-to-noise ratios (0 dB, 8 dB, and 16 dB). Effects of syllable position, signal-to-noise ratio, and articulatory features on vowel and consonant identification are discussed. The results constitute the largest source of data that is currently available on phoneme confusion patterns of American English phonemes by native listeners.

Abstract

This study investigated the perception of American English phonemes by native listeners. Listeners identified either the consonant or the vowel in all possible English CV and VC syllables. The syllables were embedded in multispeaker babble at three signalto-noise ratios (0 dB, 8 dB, and 16 dB). Effects of syllable position, signal-to-noise ratio, and articulatory features on vowel and consonant identification are discussed. The results constitute the largest source of data that is currently available on phoneme confusion patterns of American English phonemes by native listeners.