Joachim Zinke

The influence of speech coding algorithms on automatic speech recognition

Examines the influence of different coders in the range from 64 kbit/sec to 4.8 kbit/sec on both a speaker independent isolated word recognizer and a speaker verification system. Applying systems trained with 64 kbit/sec to e.g. the 4.8 kbit/sec data increases the error rate of the word recognizer by a factor of three. For rates below 13 kbit/sec the speaker verification is more affected than the word recognition. The performance improves significantly if word models are provided for the individual coding conditions. Therefore, the authors use a Gaussian classifier for estimation of the coding condition of a test utterance. The combination of this classifier and coder specific word models yields a high overall recognition performance.<<ETX>>

Extending the vocabulary of a speaker independent recognition system

The authors discuss the extension and adaptation of a speaker-independent, small-vocabulary, isolated word recognition system based on tied density hidden Markov models. In the proposed approach, the density functions are trained from a basic set of words using acoustic segmentation, position-dependent segment labeling, and clustering of the segment specific densities. Then the parameters of the word models are estimated by means of a Viterbi update procedure. With a given set of densities the Viterbi update can also be used to generate models for words not included in the basic set. The dependency between the recognition performance and the amount of reference data both for speaker-independent and speaker-dependent experiments is examined in detail. The authors compare different algorithms to avoid zero probabilities in the word models due to insufficient data.<<ETX>>

Comparison of whole word and subword modeling techniques for speaker verification with limited training data

In this paper we use whole word and subword hidden Markov models for text dependent speaker verification. In this application usually only a small amount of training data is available for each model. In order to cope with this limitation we propose a intermediate functional representation of the training data allowing the robust initialization of the models. This new approach is tested with two databases and is compared both with standard training techniques and the dynamic time warp method. Secondly, we give results for two types of subword units. The scores of these units are combined in two different ways to obtain word error rates.