Dieter Mergel

A data-driven organization of the dynamic programming beam search for continuous speech recognition

This paper describes a data-driven organization of the dynamic programming beam search for large vocabulary, continuous speech recognition. This organization can be viewed as an extension of the one-pass dynamic programming algorithm for connected word recognition. In continuous speech recognition we are faced with a huge search space, and search hypotheses have to be formed at the 10-ms level. The organization of the search presented has the following characteristics. Its computational cost is proportional only to the number of hypotheses actually generated and is independent of the overall size of the potential search space. There is no limit on the number of word hypotheses, there is only a limit to the overall number of hypotheses due to memory constraints. The implementation of the search has been studied and tested on a continuous speech data base comprising 20672 words.

Data driven search organization for continuous speech recognition

The authors describe an architecture and search organization for continuous speech recognition. The recognition module is part of the Siemens-Philips-Ipo project on continuous speech recognition and understanding (SPICOS) system for the understanding of database queries spoken in natural language. The goal of this project is a man-machine dialogue system that is able to understand fluently spoken German sentences and thus to provide voice access to a database. The recognition strategy is based on Bayes decision rule and attempts to find the best interpretation of the input speech data in terms of knowledge sources such as a language model, pronunciation lexicon, and inventory of subword units. The implementation of the search has been tested on a continuous speech database comprising up to 4000 words for each of several speakers. The efficiency and robustness of the search organization have been checked and evaluated along many dimensions, such as different speakers, phoneme models, and language models. >

Construction of language models for spoken database queries

In this paper we describe the construction of language models (finite-state networks) for a question-answering system handling database queries via spoken input(German), and report on sentence-understanding experiments with them. The world model representing the content of the permitted questions is given by a semantic structure with 5 object classes, 3 additional attribute classes and 14 relations between them The sentence structures (6 main types) are organized as a regular grammar. Acoustic recognition experiments show that the perplexity rather than the number of transitions determines the search expenditure and the error rate.

Magnetic and magneto‐optical properties of amorphous rare‐earth–transition‐metal alloys containing Pr, Nd, Fe, Co

Amorphous rare‐earth–transition‐metal alloys of composition RE1−xFex and RE1−xCox with RE=Pr, Nd and 0<x<1 were prepared by coevaporation. Also disks for magneto‐optical recording were fabricated using TbNd‐FeCo or TbPr‐FeCo layers. The magnetization, Curie temperature, uniaxial anisotropy, coercivity, and Faraday and Kerr rotation were investigated as a function of composition and temperature. The spectral variation of the Kerr rotation was measured for some alloys. The magnetization data indicate a strong dispersion of the Pr and Nd subnetworks and of the Fe subnetwork for Fe‐rich alloys. The latter gives rise to a maximum of the Curie temperature TC around x=0.7 and to very low TC values for Fe‐based alloys with x≳0.9. The magneto‐optical effects result from both the light rare earths and the transition‐metal alloys giving rise to significantly larger rotations in the visible as compared to the alloys containing heavy rare earths. Recording experiments on standard disks with a TbNd‐FeCo layer indeed re...

Phonetically guided clustering for isolated word recognition

A variant of the Markov source modelling of entire words based on automatically determined subword units is described. Each word of the vocabulary is modelled as a linear sequence of phoneme segments given by a phonetic transcription. For every phoneme a minimum and maximum duration are to be specified. Matching an utterance to the models must be performed within these absolute durational constraints. This is achieved by a dynamic programming time alignment different from the conventional ones. The acoustic emission is defined by means of phonetically labelled prototype vectors. The parameters of the models are automatically trained by an iterative procedure similar to the Viterbi algorithm. The method is applied to speaker-dependent and independent recognition of the German digits (telephone speech).

Magnetic reversal processes in exchange-coupled double layers

The classical theory to determine the interface wall energy from magnetization reversal experiments in exchange‐coupled double layers is reviewed and extended. From the energy balance three elementary reversal processes of one of the layers together with transition diagrams and corresponding reversal curves of ferrimagnetic A and P double layers are systematically derived. The wall energy can be determined from the minor loops of double layers without requiring parameters obtained on the individual layers. A novel type of partial magnetic reversal experiments is applied to amorphous rare‐earth transition‐metal bilayers. The results can be interpreted by assuming that the interface wall can move vertically through one layer and that common domains in both layers as well as separate domains may be nucleated. This can lead to asymmetric reversal loops, which make the evaluation of the wall energy density and the coercive force incorrect by the classical theory. Other asymmetries originate from the magnetic‐f...

Magnetic interface walls under applied magnetic fields

In experiments with exchange‐coupled double layers (ECDL) consisting of two ferrimagnetic layers of different composition of the general system GdTb–FeCo, it is observed that the interface wall changes its position under external fields. Furthermore, it was found that the interface wall energy σ w determined on the same ECDL can be markedly different, depending on which of the two layers’ reversal curves it was determined. It is hypothesized that this effect is due to the different external fields at which the experiments have been performed. In order to check this hypothesis, a series of bilayers with different thicknesses of the ‘‘soft’’ layer was prepared, so that the reversal loop from which σ w is determined occurs at different fields. The theoretical dependence of the interface‐wall energy on the external field as obtained from computer simulations is found to be weaker than the experimental one.

Domain formation in exchange‐coupled double layers

Exchange‐coupled double layers consisting of two ferrimagnetic layers of different composition of the general system GdTb‐FeCo were experimentally studied by magnetic reversal experiments using the Faraday effect. A novel type of magnetic reversal experiment was developed in which one of the layers is in a partly demagnetized state when the other layer is reversed. In order to interpret the resulting complicated Faraday hysteresis curves, phenomenological domain formation models are developed, including the following processes which are not possible in single layers: creation/annihilation of interface wall; reversal of the magnetization against (antiparallel to) the external field; vertical motion of the interface wall; competing reversal processes in the two layers. A method is described to estimate the distribution of compensation temperatures in a layer from reversal curves of double layers.

On the automatic training of phonetic units for word recognition

In this correspondence, we present some preliminary results on using phonetic subword units in word recognition as compared to whole word templates. The phonetic subword units are specified as either phonelike units with and without temporal structure or as diphonelike units. The determination of these subword units requires segmentation, labeling, and parameter estimation at the same time, and is performed by an iterative two-stage algorithm consisting of nonlinear time alignment and parameter estimation. Experiments were carried out, using a connected digit recognition task, to study the usefulness of the subword unit representation and the effect on recognition performance of some versions of the subword specification. The best error rates for subword units are still, by a factor of 2 or more, larger than those for whole word templates.

Bloch wall energy in (Tb,Nd)‐(Fe,Co) thin films

The Bloch wall energy has been determined for (Tb,Nd)‐(Fe,Co) magneto‐optic disks with varying Nd content. The procedure consists of applying a magnetic field to thermomagnetically written cylindrical domains and measuring the magnetic fields for which the domains expand and collapse. The wall energy can then be calculated in terms of a bubble model where a regular wall motion is assumed. The measured Bloch wall energy decreases continuously with increasing Nd content, reaching about one third of its initial value for more than 15 at. % Nd. This effect cannot be explained by a corresponding decrease in the exchange stiffness constant A or the anisotropy constant Ku. We therefore conclude that the addition of Nd, with its tendency to a sperimagnetic coupling, disturbs the magnetic ground state to such an extent that the conventional formula for the Bloch wall energy is no longer valid.