Peter Regel-Brietzmann

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.

Robust Recognition of Spontaneous Speech

This contribution describes the challenges and the progress which have been made in Verbmobil concerning robustness of speech recognition for various types of adverse conditions, like channel distortion, environmental noise and various speaker and speaking conditions. For the channel and noise problem classical approaches like cepstral bias normalization and spectral subtraction methods have been improved as well as new methods like parallel model combination. One major result is the fact, that an intelligent combination of various methods achieves the best results. Considerable progresses have also been made in research on unsupervised speaker adaptation. Several different main approaches are presented to improve robustness against variations of speaking rate, speaking style and speaker characteristics. The methods described include new estimation of the parameters for vocal tract length normalization, features and codebook transformation methods using ML algorithms, and pronunciation adaptation of the words in the lexicon.

On the performance of polynomial and HMM whole-word classifiers for digit recognition over telephone

Two statistical whole-word classification methods for speaker-independent recognition of isolated words over the telephone are compared: a mean square polynomial classifier (MSPC) and a hidden-Markov model classifier (HMMC). Both classifiers were trained and tested using a database of several hundred speakers, which was compiled for several application projects under conditions ranging from studio recording to telephone lines. In addition to comparing these two classification methods, the authors investigate whether a classifier insensitive to telephone conditions can be derived from studio training samples.<<ETX>>

Optimierung eines HMM-Spracherkennungssystems

Dieser Beitrag beschreibt mehrere Optimierungsschritte eines auf Hidden-Markov-Modellen basierenden Spracherkenners. Im einzelnen betrifft dies: Wortuntereinheiten, dynamische Merkmale, Vektorquantisierung sowie Grose und Art der verwendeten Co-debucher. Auserdem wird im Detail auf ein Verfahren zur schnellen Sprecheradaption eingegangen. Wir beschreiben dabei die Kombination von „Sprecheradaption durch Merkmalstransformation“ mit semi-kontinuierlichen Hidden-Markov-Modellen SCHMM [1, 5, 9, 10]. Da in einem solchen Erkennungssystem die Sprache eines Referenzsprechers nicht explizit in Form von Merkmals Vektoren, sondern nur in Form mehrdimensionaler Normalverteilungen vorliegt, mussen diese Verteilungen an Stelle der Merkmalsvektoren transformiert werden.