Alfred Kaltenmeier

Sophisticated topology of hidden Markov models for cursive script recognition

The paper describes an adaptation of hidden Markov models (HMM) to automatic recognition of unrestricted handwritten words. Many interesting details of a 50,000 vocabulary recognition system for US city names are described. This system includes feature extraction, classification, estimation of model parameters, and word recognition. The feature extraction module transforms a binary image to a sequence of feature vectors. The classification module consists of a transformation based on linear discriminant analysis and Gaussian soft-decision vector quantizers which transform feature vectors into sets of symbols and associated likelihoods. Symbols and likelihoods form the input to both HMM training and recognition. HMM training performed in several successive steps requires only a small amount of gestalt labeled data on the level of characters for initialization. HMM recognition based on the Viterbi algorithm runs on subsets of the whole vocabulary.<<ETX>>

Fast speaker adaptation for speech recognition systems

Different speaker adaptation methods for speech recognition systems adapting automatically to new and unknown speakers in a short training phase are discussed. The adaptation techniques aim at transformations of feature vectors, optimized with respect to some constraints. Two different adaptation strategies are discussed. The first one is based on least mean-squared-error optimization. The second method is a codebook-driven feature transformation. Both adaptation techniques are incorporated into two different recognition systems: dynamic time warping (DTW) and hidden Markov modeling (HMM). The results show that in both systems speaker-adaptive error rates are close to speaker-dependent error rates. In the best case the mean error rate of four test speakers decreases by a factor of six compared to the interspeaker error rate without adaptation. A hardware realization of the speaker-adaptive HMM-recognizer is described.<<ETX>>

DP-based wordgraph pruning

We present an efficient technique of generating word graphs in a continuous speech recognition system. The word graph is constructed in two stages. In the first stage, a huge word graph is generated as a by-product of a beam-driven forward search. Based on a dynamic-programming (DP) method, this huge word graph will be pruned in the second stage using higher level knowledge, such as n-gram language models. In this pruning stage an edge is removed if there is no path going through this edge which is better scored as the best-scored path in the word graph. The proposed technique is evaluated in the German VERBMOBIL task.

Reject management in a handwriting recognition system

The most scientific papers dealing with handwriting recognition systems make statements relating to the recognition performance based on a forced-recognition rate. This rate describes the ratio between the number of the correct recognized samples and the number of all possible samples. For systems applied in real applications this rate is not very relevant. They have to work with a very low error-rate, which can be only achieved by introducing effective reject criteria. So the real interesting thing is a function describing the recognition rate in relation to a specific error rate, including implicitly a corresponding reject rate. This paper presents two approaches for handling rejects in a hidden Markov based handwriting recognition system. The features to determine a reject are values which are derived from the hidden Markov recognizer. One of the techniques relies on relative frequencies of those values, the other one utilizes standard classification techniques to train a reject decision unit, the reject classifier. Both methods are presented with some noteworthy results.

Soft-decision vector quantization based on the Dempster/Shafer theory

The authors describe an algorithm for soft-decision vector quantization (SVQ) implemented in the acoustic front-end of a large-vocabulary speech recognizer based on discrete density HMMs (hidden Markov models) of small phonetic units. In contrast to hard-decision vector quantization (HVQ), the proposed approach transforms a feature vector into a number of symbols associated with credibility values computed according to statistical models of distances and evidential reasoning. SVQ is related to semi-continuous density HMMs (SCHMMs). In contrast to SCHMM, which is based on multidimensional, class-specific distributions of feature vectors, SVQ is based on one-dimensional distributions of distances and is therefore much simpler. Credibilities and associated symbols form the inputs to both the HMM-training and the recognition modules of the system. SVQ improves recognition results remarkably.<<ETX>>

Implementation of various LPC algorithms using commercial digital signal processors

An LPC vocoder which is compatible to the government standard algorithm LPC-10 can be implemented using three µPD 7720 Signal Peripheral Interfaces (SPI) from NEC, a microprocessor (CPU), and some other commercial integrated circuits. The main tasks, LPC analysis, pitch analysis, and synthesis, can be performed with one SPI each. Due to their large RAM storage requirements, the block-form LPC-10 algorithm is not well suited for such an implementation. Its components must be modified or even totally replaced by other algorithms which can be performed on a sample-by-sample basis and therefore require less RAM storage. In this paper some results of investigations on implementing various LPC algorithms are presented. The components of the LPC-10 algorithm are compatible with the components of the algorithms suggested in this paper, i. e. the analysis and synthesis components may be mixed with no audible degradation of speech.

A comparison of Gaussian distribution and polynomial classifiers in a hidden Markov model based system for the recognition of cursive script

Handwriting recognition systems based on hidden Markov models commonly use a vector quantizer to get the required symbol sequence. In order to get better recognition rates semi-continuous hidden Markov models have been applied. Those recognizers need a soft vector quantizer which superimposes a statistical distribution for symbol generation. In general, Gaussian distributions are applied. A disadvantage of this technique is the assumption of a specific distribution. No proof can be given whether this presupposition holds in practice. Therefore, the application of a method which employs no model of a distribution may achieve some improvements. The paper presents the employment of a polynomial classifier as a replacement of a Gaussian classifier in the handwriting recognition system. The replacement improves the recognition rate significantly, as the results show.

Handwriting Recognition by Statistical Methods

In January 1992 a project was started which is focused on the recognition of handwritten words, constraint by a given lexicon. The target application is the recognition of US city names in address reading systems.

The Recognition Algorithms

Subtask 2.1 of the P26 project was devoted to the study of the problems related to the development of the front-end of a speech understanding system. In the early stages of the project it was decided to separate the front-end, referred to in the following as the recognition module, from the understanding module, that deals with syntax and semantics. This decision was drawn taking into account several considerations mainly based on a practical point of view: the research groups working on Subtask 2.1 were at their first experience with speech understanding systems and their background was mainly in developing systems for small-vocabulary isolated and connected word recognition. Approaching the speech understanding problem required a strong effort both in knowledge acquisition and software development. For instance, methodologies for dealing with phonetic transcriptions of lexical items had to be developed from the beginning. More important was the lack of any practical feeling about the problem. Nobody knew (and very few in the world did at that time) what performance could be realistically achieved using a 1000-word vocabulary with a system based on sub-word unit modeling, hence which integrated strategy should be planned to attain a reasonably good understanding of the spoken sentences. The choice of a two-module system with a one-way interaction seemed the most appropriate for starting to acquire the proper knowledge on the problem. Besides, as people working on the two modules belonged to different groups and used different techniques as well as different programming languages (stochastic modeling and FORTRAN for the reeognition group, knowledge-based parsing and LISP for the understanding group), the best solution looked like the one by which the development of the two modules did not have to suffer from unavoidable mutual time dependencies.

Natürliche Sprache - ein multimedialer Träger von Information InfoPort - ein Projekt zur Überbrückung von Medienbrüchen bei der Verarbeitung sprachlicher Information

Es wird ein Projekt vorgestellt, das zum Ziel eine medienunabhangige Verarbeitung sprachlicher Information hat. Sprachliche Information erscheint in geschriebener oder gesprochener Form (Medien: Papier, Fax, elektonischer Text, e-mail, voice-mail, Telefon,…). Die Einsatzgebiete sind Retrieval, aktive Informationsvermittlung und Assistenz. Bereits realisierte Anwendungen liegen in den Bereichen Analyse von schriftlichen Anfragen (Geschaftsberichte), telefonische Auskunftssysteme und Datenbankzugriff (STORM). Die eingesetzten Techniken sind einerseits signalnahe Mustererkennungsalgorithmen zum Hypothetisieren von Wortern aus Bildern oder dem Sprachsignal (Dokumentbild- analyse, OCR, HMM), anderseits wissensbasierte Techniken zur Interpretation der sprachlichen Information. Eine robuste Verarbeitung verlangt eine enge Verzahnimg von Erkennung und Interpretation. Auch eine bruchteilhafte Erkennung mus interpretiert werden.