Using very large speech corpora, we can study rare but systematic pronunciation patterns in spontaneous speech. Previous studies have established that word-final alveolar consonants in English (/t/, /d/, /n/, /s/ and /z/) vary their place of articulation to match a following word-initial consonant, e.g. ran quickly → ra[ŋ] quickly. Assimilation of bilabial or velar nasals, e.g. alar[ŋ] clock for alarm clock, is unexpected according to linguistic frameworks such as underspecification theory. The existence of systematic counterexamples would challenge that theory, but these might have been previously overlooked because they are infrequent. From the c. 8-million word Audio BNC (http://www.phon.ox.ac.uk/AudioBNC) we extracted over 14,000 tokens of relevant word pairs, to determine whether non-alveolar assimilations occur and with what distribution. Word and segment boundaries were obtained by forced alignment, and F1-F3 formant frequencies were estimated using Praat. Formant frequencies in assimilation environments were compared to non-assimilating controls (e.g. come back vs. come down). We present evidence that velar and bilabial nasals sometimes do assimilate, though possibly less frequently than alveolars.

In natural speech, some phoneme transitions correspond to abrupt changes in the acoustic signal. Others are less clear-cut because the acoustic transition from one phoneme to the next is gradual. In this paper we determine the naturally occurring groups of phonemes (regardless of conventional phonetic categories) which show similar characteristics in such behavior. These data-driven groupings could be used in the design of decision-trees for context-dependent phoneme clustering, as used in large-vocabulary speech recognition and alignment systems, or during the design of speech databases for speech synthesis systems. We use 128 different Hidden Markov Model phoneme alignment systems and a large corpus of British English speech to assess the consistency with which different phoneme transitions can be identified. The phoneme transitions are grouped automatically so as to minimize the statistical differences in behavior between members of each group. In this way we derive two sets of phonemic classes, one for the first phoneme of each phoneme-to-phoneme transition, and another for the second. The grouping of the phonemes confirms that broad phonetic classes are a significant indicator of the accuracy with which boundaries can be identified, but there are a number of exceptions and some apparent sub-divisions and mergers of accepted phonetic classes. The automatic grouping of the second phonemes results in two singletons, /Z/ and /N/ (in SAMPA notation). Finally, statistics are presented which characterize the precision with which transitions between these automatic classes can be identified. These could provide weightings to be applied to different transitions to provide a more realistic assessment when evaluating the relative accuracies of different alignment systems.

A new approach is proposed for quantifying the degree of articulator movement within a phoneme as a single scalar value, using vocal tract images captured using dynamic MRI. This indicates the degree of physical movement of the articulators involved in speech production, rather than the acoustic consequences of that movement. We go on to show that this is a valid method for characterising the overall dynamics of the vocal tract, by demonstrating a coefficient of determination (R2) of 0.61 between it and a similarly-defined scalar measure of the acoustic dynamics of the signal. The calculation of the new measure involves production of images showing the exact location of any movement within the vocal tract, and additionally shows this information separately for the initial and final segments of each phoneme. Finally, we demonstrate that although some sounds may involve more movement of the articulators than would be expected from the dynamics of the acoustic signal, it is rare for the degree of articulation derived from the MRI data to be significantly less than expected. Index Terms: degree of articulation, phoneme duration, vocal tract shape, dynamic MRI.

It might be naïvely assumed that the performance of an Automatic Speech Recognition (ASR) system, and that of an Automatic Speech-to-Phoneme Alignment (ASPA) system using the same acoustic-phonetic models, would be closely related. However many researchers believe this relationship to be, at best weak - but this belief has not previously been tested in an objective and quantitative manner. This paper quantifies the strength of the relationship using analysis of data without reference to manually defined alignment labels. By avoiding comparison with a set of reference labels, both the ASR and the ASPA systems can be considered equivalent, removing any bias due to any difference of “opinion” between the human labeller and the automatic system.

Some phoneme boundaries correspond to abrupt changes in the acoustic signal. Others are less clear-cut because the transition from one phoneme to the next is gradual. This paper compares the phoneme boundaries identified by a large number of different alignment systems, using different signal representations and Hidden Markov Model structures. The variability of the different boundaries is analysed statistically, with the boundaries grouped in terms of the broad phonetic classes of the respective phonemes. The mutual consistency between the boundaries from the various systems is analysed to identify which classes of phoneme boundary can be identified reliably by an automatic labelling system, and which are ill-defined and ambiguous. The results presented here provide a starting point for future development of techniques for objective comparisons between systems without giving undue weight to variations in those phoneme boundaries which are inherently ambiguous. Such techniques should improve the efficiency with which new alignment and HMM training algorithms can be developed. Index Terms: phoneme alignment, labelling accuracy, mutual consistency, data-driven analysis, hidden Markov models.

This paper presents specific new techniques for coding of speech representations and a new general approach to coding for compression that directly utilizes the multidimensional nature of the input data. Many methods of speech analysis yield a two-dimensional (2-D) pattern, with time as one of the dimensions. Various such speech representations, and power spectrum sequences in particular, are shown here to be amenable to 2-D compression using specific models which take account of a large part of their structure in both dimensions. Newly developed techniques, multistep adaptive flux interpolation (MAFI) and multistep flow-based prediction (MFBP) are presented. These are able to code power spectral density (PSD) sequences of speech more completely and accurately than conventional methods. This is due to their ability to model nonstationary, but piecewise-continuous, signals, of which speech is a good example. Initially, MAFI and MFBP are applied in the time domain, then reapplied to the encoded data in the second dimension. This approach allows the coding algorithm to exploit redundancy in both dimensions, giving a significant improvement in the overall compression ratio. Furthermore, the compression may be reapplied several times. The data is further compressed with each application.

A two-dimensional discrete cosine transform (2-D DCT), often used for image coding, has been applied to sequences of speech spectra produced by the maximum likelihood method (MLM). The coded data was compressed by nearly 90%, reducing it to a size smaller than that needed to store the coefficients of a 10th order linear predictive coding (LPC) model. The DCT-encoded data was then reconstructed and tested for intelligibility. It was found that the two-dimensional DCT method was significantly more intelligible and more natural-sounding than the LPC technique.

This paper introduces a model being developed for estimating child and adolescent formant frequency values from adult data.The model approximates adult male and female pharyngeal and oral cavity lengths, and scales these along the corresponding male or female growth curve.The second and third formant frequencies are estimated directly from these scaled vocal tract dimensions.Two methods of establishing the first formant from the scaled data are discussed.Initial results obtained in the scaling of adult data to child values suggest that age, height and gender are all significant when estimating child formant frequency values.Furthermore, averaging of male and female data is found to be inappropriate since the differing growth rates of males and females imply that vocal tract dimensions cannot be linearly related.