Second International Conference on Spoken Language Processing (ICSLP'92)
Banff, Alberta, Canada
The research reported here is concerned with two analyses in which the identification of individual phonemes, defined across three European languages, and the discrimination between groups of phonemes are being pursued.
First, a set of acoustic-phonetic features are used to identify individual poly-phonemes, whose realisational properties across several languages are similar enough to be equated, and individual mono-phonemes, which are phonemes treated separately in these languages. Second, a confusion matrix is established from which it is possible to identify the confusions between and across sets of poly-and monophonemes. As a basis for this analysis, the phonemes are each modelled by a multi-dimensional Gaussian density probability function where the parameters are a set of principal components derived from the acoustic-phonetic features.
The acoustic-phonetic features are those describing place- and manner-of-articulation. They are derived by means of a Self-Organising Neural Network, which has been stimulated and calibrated to perform the non-linear transformation of a vector of speech signal cepstrum coefficients into the vector of acoustic-phonetic features. Results show a fairly good accuracy in identifying phonemes within feature-groups when sufficient data are available for training/simulation and testing. However, within these group-confusions occur mostly between neigboring phonemes which differ only in a few distinctive features.
Bibliographic reference. Dalsgaard, Paul / Andersen, Ove (1992): "Identification of mono- and poly-phonemes using acoustic-phonetic features derived by a self-organising neural network", In ICSLP-1992, 547-550.