## Fast Likelihood Computation Methods For Continuous Mixture Densities In Large Vocabulary Speech Recognition (1997)

Venue: | In Proc. of the European Conf. on Speech Communication and Technology |

Citations: | 14 - 10 self |

### BibTeX

@INPROCEEDINGS{Ortmanns97fastlikelihood,

author = {Stefan Ortmanns and Hermann Ney and Thorsten Firzlaff},

title = {Fast Likelihood Computation Methods For Continuous Mixture Densities In Large Vocabulary Speech Recognition},

booktitle = {In Proc. of the European Conf. on Speech Communication and Technology},

year = {1997},

pages = {139--142}

}

### Years of Citing Articles

### OpenURL

### Abstract

This paper studies algorithms for reducing the computational effort of the mixture density calculations in HMM-based speech recognition systems. These likelihood calculations take about 70 \Gamma 85% of the total recognition time in the RWTH system for large vocabulary continuous speech recognition. To reduce the computational cost of the likelihood calculations, we investigate several space partitioning methods. A detailed comparison of these techniques is given on the North American Business Corpus (NAB'94) for a 20 000word task. As a result, the so-called projection search algorithm in combination with the VQ method reduces the cost of likelihood computation by a factor of about 8 with no significant loss in the word recognition accuracy. 1.

### Citations

1115 |
Multidimensional binary search trees used for associative searching
- Bentley
- 1975
(Show Context)
Citation Context ...icient likelihood computation techniques which are based on space partitioning techniques. In particular, we describe a fast likelihood computation algorithm using the kdimensional binary search tree =-=[1, 6, 7]-=-. In addition, we present a fast log-likelihood calculation technique which is similar to the nearest neighbor search method as described in [9]. Unlike the k-dimensional binary search tree method, th... |

585 | An algorithm for finding best matches in logarithmic expected time
- Friedman, Bentley, et al.
- 1977
(Show Context)
Citation Context ...icient likelihood computation techniques which are based on space partitioning techniques. In particular, we describe a fast likelihood computation algorithm using the kdimensional binary search tree =-=[1, 6, 7]-=-. In addition, we present a fast log-likelihood calculation technique which is similar to the nearest neighbor search method as described in [9]. Unlike the k-dimensional binary search tree method, th... |

51 |
Vector quantization for the efficient computation of continuous density likelihoods
- Bocchieri
- 1993
(Show Context)
Citation Context ...search technique with two wellknown fast log-likelihood computation methods, namely the Hamming distance approximation (HDA) [2] and a vector quantization (VQ) method for mixture density preselection =-=[4, 8]-=-. The organization of this paper is as follows. In Section 2, we briefly describe the task of log-likelihood calculations using Laplacian mixture densities. In Section 3, we review the fast log-likeli... |

29 | Look-ahead techniques for fast beam search
- Ortmanns, Ney
- 2000
(Show Context)
Citation Context ...is based on a tree-organized pronunciation lexicon in connection with a bigram language model. For efficiency reasons, look-ahead pruning techniques are used during the search process as described in =-=[11]-=-. In this paper, we present two efficient likelihood computation techniques which are based on space partitioning techniques. In particular, we describe a fast likelihood computation algorithm using t... |

23 | The bucket box intersection (BBI) algorithm for fast approximative evaluation of diagonal mixture gaussians
- Fritsch, Rogina
- 1996
(Show Context)
Citation Context ...icient likelihood computation techniques which are based on space partitioning techniques. In particular, we describe a fast likelihood computation algorithm using the kdimensional binary search tree =-=[1, 6, 7]-=-. In addition, we present a fast log-likelihood calculation technique which is similar to the nearest neighbor search method as described in [9]. Unlike the k-dimensional binary search tree method, th... |

23 | State based Gaussian selection In large vocabulary continuous speech recognition using HMMs
- Gales, Knill, et al.
- 1999
(Show Context)
Citation Context ...search technique with two wellknown fast log-likelihood computation methods, namely the Hamming distance approximation (HDA) [2] and a vector quantization (VQ) method for mixture density preselection =-=[4, 8]-=-. The organization of this paper is as follows. In Section 2, we briefly describe the task of log-likelihood calculations using Laplacian mixture densities. In Section 3, we review the fast log-likeli... |

19 |
Acoustic Modeling of Phoneme Units for Continuous Speech Recognition
- Ney
- 1990
(Show Context)
Citation Context ...1 p(ljs) \Delta p(xjs; l) ; where the term p(ljs) denotes the mixture weight of the l th mixture density component. When using the maximum approximation in connection with Laplacian mixture densities =-=[10]-=-, the negative log-likelihood for a given observation vector x 2 IR D is given by: \Gammalog p(xjs) = min l ( \Gamma log p(ljs) + D X d=1 jx d \Gammaslsd j oe lsd + D X d=1 log (2 oe lsd ) ) : wheresl... |

14 | Ney: â€˜Implementation of Word Based Statistical Language Models
- Wessel, Ortmanns, et al.
- 1997
(Show Context)
Citation Context ...ment corpus comprising 310 sentences with a total of 7387 words spoken by 10 male and 10 female speakers. We used a 20 000-word vocabulary and a bigram language model with a perplexity (PP ) of 198.4 =-=[12]-=-. 199 of the spoken words were out-of-vocabulary words. The training of the emission probability distributions of the underlying hidden Markov models was performed on the WSJ0 and WSJ1 training data a... |

11 |
Hamming distance approximation for a fast log-likelihood computation for mixture densities
- Beyerlein, Ullrich
- 1995
(Show Context)
Citation Context ...likelihood computation methods. To this purpose, we combine the projection search technique with two wellknown fast log-likelihood computation methods, namely the Hamming distance approximation (HDA) =-=[2]-=- and a vector quantization (VQ) method for mixture density preselection [4, 8]. The organization of this paper is as follows. In Section 2, we briefly describe the task of log-likelihood calculations ... |

10 | Closest Point Search in High Dimensions
- Nene, Nayar
- 1996
(Show Context)
Citation Context ...gorithm using the kdimensional binary search tree [1, 6, 7]. In addition, we present a fast log-likelihood calculation technique which is similar to the nearest neighbor search method as described in =-=[9]-=-. Unlike the k-dimensional binary search tree method, this method is based on dynamic partitioning of the search space. The basic idea of the so-called projection search technique is to find all proto... |