Short Utterances Research Articles

Speaker model selection based on the Bayesian information criterion applied to unsupervised speaker indexing

In conventional speaker recognition tasks, the amount of training data is almost the same for each speaker, and the speaker model structure is uniform and specified manually according to the nature of the task and the available size of the training data. In real-world speech data such as telephone conversations and meetings, however, serious problems arise in applying a uniform model because variations in the utterance durations of speakers are large, with numerous short utterances. We therefore propose a flexible framework in which an optimal speaker model (GMM or VQ) is automatically selected based on the Bayesian Information Criterion (BIC) according to the amount of training data available. The framework makes it possible to use a discrete model when the data is sparse, and to seamlessly switch to a continuous model after a large amount of data is obtained. The proposed framework was implemented in unsupervised speaker indexing of a discussion audio. For a real discussion archive with a total duration of 10 hours, we demonstrate that the proposed method has higher indexing performance than that of conventional methods. The speaker index is also used to adapt a speaker-independent acoustic model to each participant for automatic transcription of the discussion. We demonstrate that speaker indexing with our method is sufficiently accurate for adaptation of the acoustic model.

Infant-directed speech: Final syllable lengthening and rate of speech

Speech rate has been reported to be slower in infant-directed speech (IDS) than in adult-directed speech (ADS). Studies have also found phrase-final lengthening to be more exaggerated in IDS compared with ADS. In our study we asked whether the observed overall slower rate of IDS is due to exaggerated utterance-final syllable lengthening. Two mothers of preverbal English-learning infants each participated in two recording sessions, one with her child, and another with an adult friend. The results showed an overall slower rate in IDS compared to ADS. However, when utterance-final syllables were excluded from the calculation, the speech rate in IDS and ADS did not differ significantly. The duration of utterance-final syllables differed significantly for IDS versus ADS. Thus, the overall slower rate of IDS was due to the extra-long final syllable occurring in relatively short utterances. The comparable pre-final speech rate for IDS and ADS further accentuates the final syllable lengthening in IDS. As utterances in IDS are typically phrases or clauses, the particularly strong final-lengthening cue could potentially facilitate infants’ segmentation of these syntactic units. These findings are consistent with the existing evidence that pre-boundary lengthening is important in the processing of major syntactic units in English-learning infants.

Statistical clustering and the contents of the infant vocabulary

Infants parse speech into word-sized units according to biases that develop in the first year. One bias, present before the age of 7 months, is to cluster syllables that tend to co-occur. The present computational research demonstrates that this statistical clustering bias could lead to the extraction of speech sequences that are actual words, rather than missegmentations. In English and Dutch, these word-forms exhibit the strong–weak (trochaic) pattern that guides lexical segmentation after 8 months, suggesting that the trochaic parsing bias is learned as a generalization from statistically extracted bisyllables, and not via attention to short utterances or to high-frequency bisyllables. Extracted word-forms come from various syntactic classes, and exhibit distributional characteristics enabling rudimentary sorting of words into syntactic categories. The results highlight the importance of infants’ first year in language learning: though they may know the meanings of very few words, infants are well on their way to building a vocabulary.

Robust Text-independent Speaker Recognition with Short Utterance in Noisy Environment Using SVD as a Matching Measure

A new technique for text-independent speaker recognition for noisy speech is presented. This technique is based on finding the ratio of the singular values of the feature vectors of the unknown speaker and each of the N reference features stored in the constructed database. The ith reference feature that gives the largest ratio is considered the feature of the unknown speaker.An overall correct recognition accuracy of 94% for clean speech and 32% for noisy speech of 0 dB SNR was obtained. A further step was conducted to enhance the noisy features by series expansion. The improvement in the recognition rate using the proposed SVD-based algorithm is compared with other distance measure algorithms. It is found that the proposed technique when cepstral features are used outperforms the conventional matching measure such as the Euclidean, the Weighted and the Mahalonobis distances, respectively.

Teaching receptive language to children with autism: A selective overview.

Receptive (to act based on an auditory stimulus) is an important and necessary foundational skill for children with autism. Several strategies establishing this repertoire have been developed within the field of early intensive behavior intervention (EIBI). An electronic literature review with the key words discrimination reveals extensive research by behavior analysts. However, has received far less attention. In this literature review we will present a brief summary of research findings describing the various behavioral strategies that have been effective in teaching children with autism and related developmental disabilities to act based on auditory stimuli whether labeling an object or following instructions. Keywords: Early and Intensive Behavioral Intervention, Receptive language, training techniques. Language Intervention for Children with Autism It is a well-known fact that approximately half of the autistic population fails to develop speech, and the majority fails to use speech in a functional manner when it does develop (Volkmar, 1991). For those children who do develop speech, their ability to communicate is often less advanced than their peers. These children may display speech that is often characterized by short utterances, lack of appropriate use of I/you pronouns, here/there constructions, and now/then relations, idiosyncratic use of words, and restricted use of grammatical structures (Tager-Flusberg, 1989). Furthermore, comprehension may also be severely impaired (Waterhouse & Fein, 1982). The multitude of research clearly documents that despite some recognizable patterns in the manifestation of autism there is much heterogeneity (Green, Fein, Joy, & Waterhouse, 1995). This heterogeneity is characteristic in both the symptoms that children with autism display as well as in the range of success that various children experience as a result of different types of interventions. Nevertheless, it is clear that when children experience severe delay in vocally or visually mediated their chances for inclusion educationally and socially become considerably lessened (Lovaas et al., 2003). For these reasons practitioners and researchers in the field of behavior analysis have invested their time in multiple efforts to develop based interventions that will effectively ameliorate delays in children with autism. Over the past four decades behaviorists have explored strategies for teaching to children with autism that have yielded effective data based interventions. These interventions have been described in sufficient detail and thus, have been often replicated successfully. Furthermore, these interventions have generated advanced use of in individuals with autism (Lovaas, 1987, 1993; McEachin, Smith, & Lovaas, 1993). More recently the behavioral approach referred to as Early Intensive Behavioral Intervention (EIBI) focuses on interventions for children with autism. This approach consists of intensive intervention and comprehensive curriculum protocols combining operant procedures and more recently curriculum analysis and development procedures. Specifically, EIBI focuses on simultaneously building several skill domains, including functional communication, imitation, matching-to-sample, affiliate behavior, and basic receptive and with a gradual progression to increasingly more complex domains (Pelios & Lund, 2001). Language Redefined Behaviorally The terms expressive and language are borrowed from fields in psychology other than behavior analysis and despite their general use by behavior analysts are not considered behavioral terms. A behavioral definition of receptive attempts to specify the environmental antecedents and consequences under which such responses are likely to occur. Skinner first presented a complete behavioral interpretation of in his book Verbal Behavior published in 1957. …

Event-related fMRI for the suppression of speech-associated artifacts in stuttering

The purpose of this study was to establish functional magnetic resonance imaging (fMRI) for the investigation of brain function during overt speech production in stuttering. Up to now this technique has rarely been used for the investigation of speech production paradigms because artifacts related to overt speaking largely impair the sensitivity toward task-related activation. Recently, the temporal delay of the hemodynamic response has been exploited to achieve a suppression of speech-related artifacts. By the limitation to very short utterances (one word), a temporal segregation of the respective effects was accomplished by means of an event-related experimental design. However, the investigation of speech production in persons who stutter requires a more extensive speaking situation. Since longer and more complex utterances evoke more symptoms of stuttering than reading of single words, a useful task should at least include the reading of full sentences. In this study we performed simulations to investigate the correlation of speech-related artifacts with the respective hemodynamic response in dependency on speech duration and rate of data sampling. Furthermore, we show that prolonged stimulus durations and repetition times of 3 s still allow an effective suppression of speech-related artifacts in fluent as well as in nonfluent speakers. Not only were obvious false activations at high contrast cerebrospinal fluid tissue borders widely eliminated, subjects also displayed consistent activation in speech-related and motor areas. As these results widely resemble those obtained by earlier neuroimaging studies on language production, event-related fMRI seems to be capable of recording neurophysiological correlates of overt speech production.

The production and recognition of emotions in speech: features and algorithms

This paper presents algorithms that allow a robot to express its emotions by modulating the intonation of its voice. They are very simple and efficiently provide life-like speech thanks to the use of concatenative speech synthesis. We describe a technique which allows to continuously control both the age of a synthetic voice and the quantity of emotions that are expressed. Also, we present the first large-scale data mining experiment about the automatic recognition of basic emotions in informal everyday short utterances. We focus on the speaker-dependent problem. We compare a large set of machine learning algorithms, ranging from neural networks, Support Vector Machines or decision trees, together with 200 features, using a large database of several thousands examples. We show that the difference of performance among learning schemes can be substantial, and that some features which were previously unexplored are of crucial importance. An optimal feature set is derived through the use of a genetic algorithm. Finally, we explain how this study can be applied to real world situations in which very few examples are available. Furthermore, we describe a game to play with a personal robot which facilitates teaching of examples of emotional utterances in a natural and rather unconstrained manner.

ORAL AND WRITTEN COMMUNICATIVE PATTERNS IN ENGLISH LANGUAGE TEACHING (ELT) FOR THE HOTEL AND TOURISM INDUSTRY

Regarding language as more than simply a system of rules, we need to distinguish between knowing various grammatical rules and being able to use the rules effectively and appropriately when communicating. This view has underpinned communicative language teaching which is a very important component of the English teaching material used with our students. It should be emphasized that communicative language teaching in the field of the hotel and tourism industry covers spoken as well written language for specific purposes. Written language being characterised by well-formed sentences integrated into structured paragraphs differs sharply from spoken language consisting of short utterances with the loosely organised syntax, the use of non-specific words and phrases and the use of fillers such as "well", "oh" etc. The English teaching material we are working with at our associate degree faculty tries to cover the needs of oral and written communication in the field of the hotel and tourism industry and because of that some of its parts and exercises are presented in this paper.

High speed MRI of laryngeal gestures during speech production

Dynamic sequences of magnetic resonance images (MRI) of the vocal tract were obtained with a frame rate of 144 frames/second. Changes in vertical position and length of the vocal folds, both observable in the mid-sagittal plane, have been argued to play a role in consonant production in addition to their primary function in the control of vocal fundamental frequency (F0) [W. G. Ewan and R. Krones, J. Phonet. 2, 327–335 (1974); A. Lofqvist et al., Haskins Lab. Status Report Speech Res., SR-97/98, pp. 25–40, 1989], but temporal resolution of available techniques has hindered direct imaging of these articulations. A novel data acquisition sequence was used to circumvent the imaging time imposed by standard MRI (typically 100–500 ms). Images were constructed by having subjects rhythmically repeat short utterances 256 times using the same F0 contour. Sixty-four lines of MR data were sampled during each repetition, at 7 millisecond increments, yielding partial raw data sets for 64 time points. After all repetitions were completed, one frame per time point was constructed by combining raw data from the corresponding time point during every repetition. Preliminary results indicate vocal fold shortening and lowering only during voiced consonants and in production of lower F0.

The effect of temporal manipulation on the perception of disfluencies as normal or stuttering

The purpose of this investigation was to study the effect of temporal features within repetition of speech segments on the perception of stuttering. Past research has provided evidence that certain temporal aspects of repetitions produced by people who stutter tend to be shorter than those produced by normally fluent speakers. The effect of these temporal factors on the perception of the disfluency as “stuttering” or “normal” has not yet been studied. Conversational speech of five children who stutter was recorded. Two short utterances, one containing part-word repetition (PWR) and one containing whole-word repetition (WWR), were identified in the speech of each child and then manipulated by the CSL and CSpeech computer softwares. Two selected elements within repetitions, namely the vowel of the repeated unit and the interval between the repeated units (e.g., but-but), were lengthened to simulate normal disfluency. Results indicated that both factors (interval duration and vowel duration) moderately affected listeners' perception. In general, repetitions with short vowel and interval durations were judged as more representative of stuttering, whereas repetitions with longer vowel and interval duration were judged as more representative of normal speech. Learning outcomes: As a result of this activity, the reader will learn about (1) various factors that influence the perception of disfluent segments as stuttering, (2) the special effect of duration of specific elements within repetitions on the perception of disfluency as stuttering, and (3) the possible implications of the new information for therapeutic considerations.

Tactical communications bring new challenges to automatic speech recognition

Automatic speaker-independent recognition of conversational speech has been making significant progress in recent years. Military communications challenge the robustness of current recognition systems. Compared to telephone conversations or news broadcasts, tactical communication has a reduced vocabulary. It consists of short utterances, limited to the task at hand, with occasional chat words. What makes recognition difficult is high levels of background noise in tactical environments (helicopters, tanks) and the degradation of the signal by military microphones, including noise canceling microphones. Of perhaps more importance, people alter their speech to overcome these degradations and the loss of intelligibility from communicating over vocoders. The DARPA-sponsored spine program is exploring these issues using speech from pairs of participants performing a collaborative task while communicating between separate sound booths. Each person sits in an accurately reproduced military background noise environment and uses the type of microphone and ear protection that would normally be found in that environment. The communication is conducted through vocoders. This presentation will review results from the spine evaluations of state-of-the-art speech recognition systems and from analyses of the transcripts in an attempt to understand how environmental factors and speaking changes affect recognizer performance. [Work supported by DARPA.]

SLI: What We Know and Why It Matters

SLI: What We Know and Why It Matters

Annotation in the SpeechDat Projects

A large set of spoken language resources (SLR) for various European languages is being compiled in several SpeechDat projects with the aim to train and test speech recognizers for voice driven services, mainly over telephone lines. This paper is focused on the annotation conventions applied for the Speechdat SLR. These SLR contain typical examples of short monologue speech utterances with simple orthographic transcriptions in a hierarchically simple annotation structure. The annotation conventions and their underlying principles are described and compared to approaches used for related SLR. The synchronization of the orthographic transcriptions with the corresponding speech files is addressed, and the impact of the selected approach for capturing specific phonological and phonetic phenomena is discussed. In the SpeechDat projects a number of tools have been developed to carry out the transcription of the speech. In this paper, a short description of these tools and their properties is provided. For all SpeechDat projects, an internal validity check of the databases and their annotations is carried out. The procedure of this validation campaign, the performed evaluations, and some of the results are presented.

Phonology and children with specific language impairment: Status of structural constraints in two languages

Two studies are reported in which the phonological characteristics of preschool-age children with specific language impairment (SLI) are compared with those seen in younger normally developing children matched for mean length of utterance and consonant inventory size. The productions of both English-speaking and Italian-speaking children with SLI were more likely to deviate from the adult standard than the productions of the younger control children. In Italian, the children with SLI had more difficulty than the younger controls in the use of non-final weak syllables; in English, the children with SLI were poorer than the younger controls in the use of non-final weak syllables, word-final consonants, and word-final consonant clusters. These are the same phonological details that are required for several grammatical inflections and many function words in the two languages. However, the children with SLI were also less consistent than their younger compatriots in using consonants in structurally simple words. These findings provide evidence for the view that for many preschool-age children with SLI, phonological problems go beyond those that might be predicted on the basis of the children's short utterances and limited consonant inventories.

Speaker identification using neural networks and wavelets.

Multiresolution decomposition and pattern-recognition techniques enable identification in noisy environments. The cocktail party effect describes the phenomenon in which humans can selectively focus attention to one sound source among competing sound sources. This is an ability that is hampered for hearing-impaired individuals. In this article, the authors present an off-line system that uses wavelets to generate multiresolution time-frequency features that characterize the speech waveform to successfully identify a speaker in the presence of competing speakers. This system is successful for short utterances and has also been applied to interspeaker speech recognition. The authors also discuss ALOPEX, which is an optimization paradigm that incorporates the above-mentioned features into a pattern-recognition system through template matching or connectivity weight updating in a feedforward artificial neural network.

Children's responses when interviewers distort details during investigative interviews

Objective. Misunderstandings sometimes occur during investigative interviews when an interviewer distorts details provided by the child. This was examined in the present study by investigating (i) when and how ‘distortions’ occur, and (ii) children's responses to the distortions.Method. In all, 140 such distortions were identified in 68 formal investigative interviews of 3‐ to 14‐year‐olds who had made allegations of sexual or physical abuse. Features of the utterance in which the distortion was made were coded (e.g. the complexity of the utterance, the topic that was being discussed) and analysis of variance and chi‐square tests were used to determine whether these affected children's responses to the interviewers' distortions.Results. One‐third of the distortions were corrected by the children. The children were most likely to correct interviewers when the distortions were embedded in short utterances and when the interviewers changed aspects of someone's identity. When children did not correct the interviewers, the interviewers continued to misrepresent the details in the remainder of the interview.Conclusions. Although some children felt comfortable correcting the interviewers, a substantial portion of the distortions were not corrected. The results of the study suggest that it might be valuable to inform children at the beginning of the interview that they can and should correct interviewers, and to provide children with the opportunity to practice doing this prior to the substantive phase of the interview.

Rapid changes in speech parameters with a change in auditory stimulation from a cochlear implant

To investigate the speed of speech production changes in response to a change in hearing, the speech processors of two cochlear implant users were switched on and off a number of times in a single experimental session for each subject. The subjects repeated short utterances (‘‘a said,’’ ‘‘a shed,’’ ‘‘a sad,’’ and ‘‘a shad’’) many times in semi-random order. The transitions between hearing (on) and nonhearing (off) states were timed to occur between utterances, and the number of utterances between transitions was varied to minimize subject anticipation of the change. Using transition times as line-up points, values of median and skewness of sibilant spectra and vowel F1, F2, duration, and SPL were each averaged over repetitions of each utterance and compared across the on–off and off–on transitions. Vowel SPL and duration had changed by the first utterance following the transition, indicating that the subjects were using subtle aspects of the processor output to detect its state even in the absence of overt sound input. Changes in spectral median and skewness and F1 and F2 were less immediate, more varied, and differed between the two subjects. The paradigm is currently being refined to further investigate this issue. [Work supported by NIH.]

MRI-based speech production study using a synchronized sampling method.

The magnetic resonance imaging (MRI) technique is primarily used to obtain still pictures of body sections, but it also provides a practical means for motion imaging. In this work, the synchronized sampling method originally developed for cardiac motion imaging was applied to the recording of articulatory movements while the subject repeated a short utterance synchronized with the scan sequence. Results showed dynamic movements ofthe speech organs in the midsagittal plane. A few key frames from the movie sequence are presented as an example of its application to speech production research.

An Efficient, Probabilistically Sound Algorithm for Segmentation andWord Discovery

This paper presents a model-based, unsupervised algorithm for recovering word boundaries in a natural-language text from which they have been deleted. The algorithm is derived from a probability model of the source that generated the text. The fundamental structure of the model is specified abstractly so that the detailed component models of phonology, word-order, and word frequency can be replaced in a modular fashion. The model yields a language-independent, prior probability distribution on all possible sequences of all possible words over a given alphabet, based on the assumption that the input was generated by concatenating words from a fixed but unknown lexicon. The model is unusual in that it treats the generation of a complete corpus, regardless of length, as a single event in the probability space. Accordingly, the algorithm does not estimate a probability distribution on wordss instead, it attempts to calculate the prior probabilities of various word sequences that could underlie the observed text. Experiments on phonemic transcripts of spontaneous speech by parents to young children suggest that our algorithm is more effective than other proposed algorithms, at least when utterance boundaries are given and the text includes a substantial number of short utterances.

Reducing the effects of linear channel distortion on continuous speech recognition

Linear channel compensation in speech recognition typically involves estimating an additive shift in the cepstral domain. This paper explores both Bayesian and maximum likelihood techniques to transform either the features or the model parameters. Experiments on the Macrophone corpus show error rate reductions of up to 16% over cepstral mean subtraction for short utterances.