Bernhard Kämmerer

Experiments for isolated-word recognition with single- and two-layer perceptrons

Abstract Several design strategies for feed-forward networks are examined within the scope of pattern classification. Single- and two-layer perceptron models are adapted for experiments in isolated-word recognition. Direct (one-step) classification as well as several hierarchical (two-step) schemes have been considered. For a vocabulary of 20 English words spoken repeatedly by 11 speakers, the word classes are found to be separable by hyperplanes in the chosen feature space. Since for speaker-dependent word recognition the underlying data base contains only a small training set, an automatic expansion of the training material improves the generalization properties of the networks. This method accounts for a wide variety of observable temporal structures for each word and gives a better overall estimate of the network parameters which leads to a recognition rate of 99.5%. For speaker-independent word recognition, a hierarchical structure with pairwise training of two-class models is superior to a single uniform network (98% average recognition rate).

Large-vocabulary isolated word recognition with fast coarse time alignment

An isolated word recognition system for large vocabularies (1000 words and up) with an average recognition rate of more than 98 per cent is presented. Each utterance is characterized by a sequence of feature vectors which are obtained by an autocorrelation analysis. The resulting coefficients are quantized in such a way, that an entire feature vector can be stored in a single data word. A distance measure adapted to this representation has been developed. The classification is performed hierarchically in two steps. In the preselection stage, the words are divided into 16 segments and hardware is employed for a coarse nonlinear mapping. A short ranked list of candidates is processed by the following final classifier which performs a nonlinear time alignment of the fully resolved patterns using Dynamic Programming. Thus a short response time and a high recognition performance of the system are achieved. Without full use of parallelism an overall response time for the large vocabulary of less than one second is obtained on a signal processor.

Special feature vector coding and appropriate distance definition developed for a speech recognition system

In this paper a single-word speech-recognition system based on autocorrelation feature vectors is presented. Existing comparable systems use an extra data word for each coefficient of a feature vector. Normally for large vocabularies, vector quantization is performed in order to reduce the resulting large amount of data. Another way to reduce the storage needed is proposed by using a rough vector coefficient quantization instead of vector quantization. If, for example, 16 autocorrelation coefficients coded with two bits each are stored in one 32 bit data word, one obtains, besides an optimal use of the available storage, a good facility of computing a distance between pairs of feature vectors in a very fast way. A modified distance measure based on the cityblock distance is introduced. It only takes about 200 ns to compute one distance with the aid of appropriate programmed read only memories. If coefficient quantization is involved instead of vector quantization and the modified distance measure is used, there is no loss of accuracy. In each case we obtained recognition rates better than 95% for a 250-word recognition system.

Verbesserung der sprecherunabhängigen Erkennung

Im folgenden werden mehrere Ansatze zur Verbesserung der sprecherunabhangigen Spracherkennung erlautert. Diese beziehen sich auf eine moglichst gute Ausnutzung des Trainings-materials mehrerer Referenzsprecher, da bei der sprecherunabhangigen Erkennung wahrend der Lexikonerstellung keine Sprachproben des neuen Sprechers fur die Gewinnung sprecherspezifischer Information vorliegen.

Sprecherunabhängigkeit und Sprecheradaption

Einem Sprecher ist es nicht moglich, zwei Auserungen eines Wortes mit identischen Sprachsignalen zu erzeugen. Somit treten selbst bei der sprecherspezifischen Erkennung unterschiedliche Muster fur ein Wort auf. Dies kann in der Erkennungsphase zu Klassifikationsfehlern fuhren. Spezifische Unterschiede zwischen Sprechern bewirken dagegen eine weitaus grosere Streuung der Merkmalcharakteristika, so das bei der Verwendung von Sprachmustern eines Sprechers als Prototypen fur die Erkennung von Sprachmustern eines zweiten Sprechers eine Zunahme der Klassifikationsfehler auftritt. Daraus ergibt sich die Notwendigkeit, fur das Ziel eines universell einsetzbaren Erkennungssystems sprecherunabhangig optimierende und sprecheradaptive Verfahren zu entwickeln und die klassischen Ansatze damit zu erganzen.

Ansätze für die Sprecheradaption

Dieses Kapitel beschreibt die untersuchten Ansatze, Wortmuster eines Sprechers so zu verandern, das sie den Wortmustern eines anderen Sprechers angepast sind und somit auch ohne sprecherspezifisches Training aller Referenzmuster eine gute Erkennungsleistung des Systems ermoglichen. Die Parameter fur diese Anpassung sollen dabei aus wenigen Auserungen gewonnen werden. Hierbei liegt die Annahme zugrunde, das von den Charakteristika der Merkmale weniger Worter auf das Verhalten der Merkmale aller Worter geschlossen werden kann. Damit diese Annahme moglichst gut erfullt wird, mus die Vorgehensweise bei der Adaption sorgfaltig gewahlt werden.