This paper presents a method to address a significant real life problem of embedded systems. Such systems are characterized by limited resources. Restrictions are placed on the amount of available memory for the acoustic models as well as a limited computational capacity to perform the needed distance calculations. Hence the trade-off between optimal performance and meeting such requirements needs a tailor-made HMM set solution. Starting from single densities, boosting the performance is normally done by blindly splitting model densities hence doubling the amount of memory for the acoustic models. The approach in this paper proposes an iterative length adaptation (ILA) scheme to change the model length to the needs of the acoustic events. It improves the model accuracy through a decrease of the overall acoustic distance score while minimizing the amount of additional memory for the acoustic model. The improved HMMs result in a decrease of the achieved WER outperforming the acoustic models utilizing two mixture densities per state whereas the number of model densities are only increased by 21% in comparison to the single density models.

Entering city names is an important issue for various speech driven applications such as telephone directory assistance. This paper proposes a system that combines word recognition with utterance verification and spelling as a fall back strategy. Word recognition experiments show that the use of a medium size vocabulary yields the lowest error rate when only permitting very few false acceptances. For letter recognition discriminatively trained letter models are applied. In order to shorten the spelling procedure two abort conditions are introduced which reduce the number of letters that have to be spelled. The system handles 96.5% of all calls correctly while less than 45% of all callers must spell the name.

The assumption of statistically independent feature vectors within the HMM approach is a well known problem. The aim of this study is to explore a simple and feasible method, that takes the correlation of adjacent feature vectors into account. A so called correlated HMM, that estimates the emission probability of a state with respect to correlated feature vectors, is built by combining two separate knowledge sources. On the one side, a traditional HMM provides an emission probability under the condition of a certain state, whereas on the other side a linear predictor delivers an emission probability considering the previous feature vectors. The efficiency of this method is shown with the help of the German SpeechDat(M) database. The application of the correlated HMM within the verification procedure of a keyword spotter provided an improvement of the figure-of-merit from 87.1% to 88.6%.

This paper refers to our prosperous development of algorithms for detecting keywords in continuous speech. Two different approaches to define confidence measures are introduced. As an advantage, these definitions are theoretically calculable without artful tuning. Moreover, two distinct decoding algorithms are presented, that incorporate these confidence measures into the search procedure. One is a new possibility of detecting keywords in continuous speech, using the standard Viterbi algorithm without modeling the non-keyword parts of the utterance. The other one is an improved further development of an algorithm described in [1], also without the need of modeling the non-keyword parts.

Keyword spotting is a very forward looking and promising branch of speech recognition. The paper presents a HMM based keyword spotting system, which works with a new algorithm. The first discussion topic is the description of the search algorithm, that needs no representation of the non keyword parts of the speech signal. For this purpose, the computation of the HMM scores and the Viterbi algorithm had to be modified. The keyword HMMs are not concatenated with other HMMs, so that there is no necessity for filler or garbage models. As a further advantage, this algorithm needs only low computational expense and storage requirement. The second discussion topic is the determination of a optimal decision threshold for each keyword. In order to decide between the two possibilities “keyword was spoken” and “keyword was not spoken”, the scores of the keywords are compared with keyword specific decision thresholds. The paper introduces a method to fix decision thresholds in advance. Starting with measured phoneme distributions, the score distributions of whole keyword models can be calculated. Furthermore, these keyword distributions form the basis of the computation of decision thresholds. Tests with spontaneous speech databases yielded 73.9% Figure Of Merit when using context dependent HMMs. The detection rate at 10 fa/kw/h comes to 80%