Search

One of the most basic comparisons between objects is to ask how similar they are. Linguistics and phonology are founded on this question. The classic definitions of phonemes and features involve contrast between minimal pairs. A minimal pair of words requires that there be two sounds that are dissimilar enough for the words to be considered different. Otherwise we wouldn't speak of a minimal pair of words but rather of a single word with two meanings. Likewise, phonetic similarity is needed to group together separate instances into a single word or sound. Without some intuition about which sounds are so similar that they should be treated as instances of the same linguistic object, the field would be no more than a collection of trillions of disconnected examples. So, it is important to have a measure of dissimilarity between sounds. Some exist already: e.g. measures of dissimilarity between the input and output of speech codecs have been used as a way of quantifying their performance (e.g. Gray, Gray, and Masuyama 1980). Cepstral distance has been frequently used, and the Itakura-Saito divergence is also widely used. But, none of these have been explicitly calibrated against the differences in speech that are important to human language. In this paper, we do so.

This describes the FIAT 1.2 file format. FIAT is a simple and flexible file format that stores multicolumn data, is readable by humans with a text editor, is readable by many programs that read CSV format, and has the ability to store metadata and arbitrary header information.

People often speak about 'rhythm' of a language and that different languages or dialects have different rhythm. This has lead to substantial interest in quantitative measures that would capture this distinction. Traditionally, most of the rhythm metrics relied on duration. However, experimental evidence suggests that the perception of rhythm rests on a combination of different coustic properties that are not limited to duration. These include changes in F0, rate of spectral change and auditory prominence. In this poster we present an overview of different approaches to measuring linguistic rhythm and report experimental results that evaluate reliability of some of these methods.

The motivation of this paper is to build a bridge between phonology and phonetics with implementation models. The challenge is to explain a wide range of phonetic forms in diverse speaking styles, including laboratory speech, spontaneous speech, fluent and non-fluent speech and model them as orderly variations of one coherent communication system. In the paper, we will focus on the effect of sentence planning on articulatory gestures. Data from spontaneous speech provides strong evidence for anticipatory effects and partial reduction effects. How and when they occur reflect the speaker's sentence planning strategies. It is hypothesized that much of the discrepancy between laboratory speech and spontaneous speech can be accounted for with a model that can represent these effects, such as the weights used in the Stem-ML model.

We use Stem-ML to build an automatic learning system for Mandarin prosody that allows us to make quantitative measurements for prosodic strengths. Stem-ML is a phenomenological model of the muscle dynamics and planning process that controls the tension of the vocal folds. Because Stem-ML describes the interactions between nearby tones and accents, we were able to use a highly constrained model with only one accent template for each lexical tone category, and a single prosodic strength per word. The model accurately reproduces the intonation of the speaker, capturing 87% of the variance of the speech's fundamental frequency, f0. The result reveals strong alternating metrical patterns in words, and suggests that the speaker uses word strength to mark a hierarchy of sentence, clause, phrase, and word boundaries.

We present the results of a quantitative study of durational variability in several British dialects. The study is based on the IViE corpus (Grabe 2001), which contains recordings of speakers from seven central urban dialects of English from the UK and Ireland. The corpus used in this study included speakers from Belfast, Cambridge, Newcastle, Leeds, Dublin, monolingual London residents of Jamaican decent and bilingual Bradford residents of Punjabi decent. We used reading and re-telling of "Cinderella" story by 6 female and 6 male speakers of each dialect (7 female and 5 male for London). The read version of Cinderella was split into 5 parts. The total total number of sound files used for analysis was 465. The corpus was automatically segmented into vocalic and consonantal intervals. We then computed 15 published measures of durational variability (for example, %V or VnPVI). To compare the variation between the dialects with variation within the dialects, we used classifier algorithms based on all possible combinations of 1, 2 or 3 measures (more than 2000 in total). We found that there is a substantial overlap between the dialects leading to a relatively low classification rate: based on various combinations of the rhythm measures, the classifiers could correctly identify the dialect of not more than 30% of data (chance=15%). We see an increase in performance of the classifiers based on greater number of rhythm measures. This shows that our previous finding that rhythm is a multidimensional phenomenon (Loukina et al. 2009) is true not only for different languages, but also for different dialects of the same language. Our classifiers showed higher identification rates for Belfast English. These results agree with the results previously reported by Ferragne (2004, 2008) who conducted a similar study on a different corpus of British dialects: they found that rhythm measures computed on Ulster English were different from other accents in their corpus. Contrary to their result, we find little overlap between Belfast and Dublin English. We also found better classification rates for the two 'ethnic minorities' accents: Punjabi and Jamaican The patterns we found based on measures of durational variability are partially similar to what has been reported for intonation contours. Based on the same corpus, Coleman, Grabe and Kochanski (submitted) reported significant differences between Belfast and other dialects. At the same time, duration-based measures do not show the distinction between 'Northern' and 'Southern' dialects observed by Coleman et al., but separate the two dialects spoken by 'ethnic minorities'. This suggests complex patterns of interaction between different aspects of prosodic structure.

The tutorial will summarize the state of the art of prosody research with an emphasis on intonation ($f_0$). We will present it in a data and knowledge driven fashion, starting with physiology and speech data and building towards theoretical descriptions. We will cover a wide range of intonation phenomenon, discussing the variations, the cause of variations, and their implications to prosodic modeling. Major concepts that will be addressed are: *** The physical bases of prosody: what gestures and acoustic features carry information, and how the acoustics relate to physiology. *** Prosody as a communications channel ** Accents and tones ** . Intonation type (Question vs. declarative ldots) ** Focus, prominence,. discourse segments, {em etc} ** Prosody used to convey emotion ** . Prosody tied to the physical system: declination . *** A survey of algorithms for modeling prosody. We will comment on the strengths and limitations of a variety of theories and modeling techniques. ** Basic Questions: * How is all this information carried on a shared channel? * How is it encoded? * How much of prosody can be predicted from text? ** Intonation Models. * Some models describe prosody as a single composite signal, others as discrete signals. ** Duration.

We present a step towards a mathematical description of style in speech and singing. We represent a style of speech or singing as a set of portable prosodic features along with a set of rules to choose where the features are to be applied. Speech and singing can both be conveniently described by separating their style from their content. In both speech and singing, style is implemented in terms of acoustic parameters like pitch and loudness. The features we use correspond to musical embellishments or prosodic accents and phrase curve in speech, and are implemented via Stem-ML (Soft TEMplate Mark-up Language). We present several examples to illustrate the usefulness of the style/content separation, and the similarity between speech and music.

In this paper, we investigate how intonation is used to confirm a word in English. This intonation type is challenging to model, as it mixes narrow focus and question with variations based on accent location, phrasing and speaking rate. We build a model that predicts the intonation from the text, using an extremely simple intonational phonology. One can interpret some of the parameters of the model as detailed description of accent shapes and others as prosodic strengths which carry phrasing information. The RMS deviation is 21 Hz or 1.7 semitones, a result comparable to machine learning methods, but with far fewer parameters that need to be learned. Furthermore, the model handles both fast and slow speech with the same set of parameters in a principled way. The model incorporates some aspects of muscle dynamics, and its ability to predict f0 at different speaking rates is confirmation that an articulatory approach to f0 modeling is appropriate.

In most of the world's languages, one can transform a statement into a question by raising the pitch at the end. This transformation is dubious in tone languages, as it could transform one lexical item into another. * How does one ask question in a tone language? * Is there is a question phrase curve? * Is there a question boundary tone? We build and train models of Mandarin Chinese intonation to answer these question. The resulting models have RMS errors of 10 Hz, or 1 semitone. We find that questions are marked by: * More careful intonation, and a greater range of speech fundamental frequency at the end of the sentence. * A slightly raised, but otherwise unremarkable, phrase curve.

We explored a database covering seven dialects of British and Irish English and three different styles of speech to find acoustic correlates of prominence. We built classifiers, trained the classifiers on human prominence/nonprominence judgments, and then evaluated how well they behaved. The classifiers operate on 452 ms windows centred on syllables, using different acoustic measures. By comparing the performance of classifiers based on different measures, we can learn how prominence is expressed in speech. Contrary to textbooks and common assumption, fundamental frequency (f0) played a minor role in distinguishing prominent syllables from the rest of the utterance. Instead, speakers primarily marked prominence with patterns of loudness and duration. Two other acoustic measures that we examined also played a minor role, comparable to f0. All dialects and speaking styles studied here share a common definition of prominence. The result is robust to differences in labeling practice and the dialect of the labeler.

Can one trust experiments conducted with repetitive speech to represent normal language behaviour? We compare the spectra of repetitive productions of phrases with the same phrases read from a randomised list. We use a data-driven spectral distortion measure that is trained to respond to linguistically relevant differences: it is based on a classifier that separates sounds into linguistically equivalent or not. We find that repetitive speech is not distinct from individually uttered speech. The difference between these two sorts of speech is smaller than variation within each. It is substantially smaller than typical differences between utterances produced by different subjects.

This poster describes the creation of an automatic word and phoneme alignment between the audio recordings of the Spoken British National Corpus (BNC) and their corresponding word-level transcriptions. The work presented here is part of the “Mining a Year of Speech” project which aim is to produce automatic speech-to-phoneme alignments of an approximately one year of audio recordings.The Spoken BNC recordings consist of unscripted, spontaneous speech conversations in different recording conditions, accents and background noises. The range of topics covers from radio programs to family conversations, council meetings or chemistry courses. The Spoken BNC was originally recorded on analogue cassette tapes between 1991 and 1994. These tapes have been recently-digitised by the British Library. The resulting dataset is composed of approximately 2,000 digital audio files with an average duration of 45 minutes and their associated word-level transcriptions. This poster describes the dataset, the automatic alignment process, the results obtained and the difficulties encountered.

This paper introduces Stem-ML, which is a model of the prosody generation process with an associated description language, and suggests how it may help prosody recognition. We applied Stem-ML modeling to three topics: the modeling of prosodic strengths, intonation types, and noun phrase patterns. Stem-ML parameters derived from F0 contours may have a more consistent relationship with prosodic events than raw F0 values. This may improve identification of accent classes, accent strengths, and intonation types.

A novel procedure is described for accelerating the convergence of Markov chain Monte Carlo computations. The algorithm uses an adaptive bootstrap technique to generate candidate steps in the Markov Chain. It is efficient for symmetric, convex probability distributions, similar to multivariate Gaussians, and it can be used for Bayesian estimation or for obtaining maximum likelihood solutions with confidence limits. As a test case, the Law of Categorical Judgment (Corrected) was fitted with the algorithm to data sets from simulated rating scale experiments. The correct parameters were recovered from practical-sized data sets simulated for Full Signal Detection Theory and its special cases of standard Signal Detection Theory and Complementary Signal Detection Theory.

"Hackers" have written malicious programs to exploit online services intended for human users. As a result, service providers need a method to tell whether a web site is being accessed by a human or a machine. We expect a parallel scenario as spoken language interfaces become common. In this paper, we describe a Reverse Turing Test (i.e., an algorithm that can distinguish between humans and computers) using speech. We present a test that depends on the fact that human recognition of distorted speech is far more robust than automatic speech recognition techniques. Our analysis of 18 different sets of distortions demonstrates that there are a variety of ways to make the problem hard for machines. In addition, humans and speech recognition systems make different kinds of mistakes, and this difference can be employed to improve discrimination.

By proper control of prosody, text-to-speech systems already have the capacity to imitate distinctive speaking styles. We show two examples where we can capture the critical features: the singing style of Dinah Shore and the speaking style of Martin Luther King Jr. The styles are described by Stem-ML tags (soft template mark-up language), which offers the flexibility needed to control accent shapes, phrasal pitch contours, and amplitude profiles, for speech as well as for singing.