## Combining Local PCA and Radial Basis Function Networks for Speaker Normalization (1995)

Venue: | Eds.), Proceedings of the 1995 IEEE Workshop on Neural Networks for Signal Processing V |

Citations: | 7 - 0 self |

### BibTeX

@INPROCEEDINGS{Furlanello95combininglocal,

author = {C. Furlanello and D. Giuliani},

title = {Combining Local PCA and Radial Basis Function Networks for Speaker Normalization},

booktitle = {Eds.), Proceedings of the 1995 IEEE Workshop on Neural Networks for Signal Processing V},

year = {1995},

pages = {233--242},

publisher = {IEEE}

}

### OpenURL

### Abstract

Complex multidimensional data may naturally require the decomposition of a regression/classification problem over local regions. Moreover, both global and local anisotropy can be present. We propose to address both problems with a flexible neural network structure embedding data quantization and coordinate transformations. The solution is applied in this paper to speaker normalization. The spectral mapping is realized as a weighted superposition of local neural mappings, estimated between subregions of a new speaker acoustic space and that of a reference speaker, combined with global and local space transformations. The local mappings are realized using the Generalized Resource Allocating Network (GRAN) model, a general RBF scheme that allows recursive allocation of kernels. The space transformations are based upon projections over the principal components, separately estimated for the global space and for the local subregions of the input and output acoustic spaces. 1 INTRODUCTION Th...

### Citations

332 | Regularization theory and neural networks architectures
- Girosi, Jones, et al.
- 1995
(Show Context)
Citation Context ...SOURCE ALLOCATING NETWORK MODEL The RBF network model applied in this paper introduces recursive allocation of elliptical kernels within the Generalized Regularization Network approximation scheme in =-=[13]-=- according to a heuristic procedure of [14]. The Generalized Resource Allocating Network (GRAN) model is thus a nonparametric kernel regression method that allows optimization of kernel coefficients a... |

208 | A theory of networks for approximation and learning
- Poggio, Girosi
- 1989
(Show Context)
Citation Context ...ine linear term and a weighted metrics is considered in the input space in the general case. The optimization of the weighted metrics is implemented, as in the generalized versions of RBF networks of =-=[15]-=-. Efficient on-line optimization of parameters is achieved by an automatic differentiation technique, which allows reuse of network evaluation for on-line optimization of parameters with respect to di... |

178 | A Resource-Allocating Network for Function Interpolation
- Platt
- 1991
(Show Context)
Citation Context ...work model applied in this paper introduces recursive allocation of elliptical kernels within the Generalized Regularization Network approximation scheme in [13] according to a heuristic procedure of =-=[14]-=-. The Generalized Resource Allocating Network (GRAN) model is thus a nonparametric kernel regression method that allows optimization of kernel coefficients and locations, of different local bandwidths... |

64 | Combining estimators using non-constant weighting functions
- Tresp, Tanguchi
- 1995
(Show Context)
Citation Context ... is an example of optimization of the weights (w l ) l=1:::L . A similar "softmax" combination of local linear function was realized in [11], with weights estimated by an RBF network, while =-=reference [12] gives ind-=-ications for a probabilistic treatment of the general case. Here the joint use of this "architectural" approach is exploited combining local PCA transformations and RBF networks. Although th... |

60 |
Jordan,“Learning Piecewise Control Strategies in a Modular Neural Network Architecture
- Jacobs, I
- 1993
(Show Context)
Citation Context ...tribution parameters initialized from the clustering procedure. The idea underlying the model is strongly related to the concept of switching regression [7, 8] and to the "mixture of expert"=-= approach [9, 10]. The Expe-=-ctation step in the EM algorithm [7, 9] is an example of optimization of the weights (w l ) l=1:::L . A similar "softmax" combination of local linear function was realized in [11], with weig... |

52 |
Switching Regression Models and Fuzzy Clustering
- Hathaway, Bezdek
- 1993
(Show Context)
Citation Context ...ch Gaussian modeling a subregion and having distribution parameters initialized from the clustering procedure. The idea underlying the model is strongly related to the concept of switching regression =-=[7, 8] and to the "mi-=-xture of expert" approach [9, 10]. The Expectation step in the EM algorithm [7, 9] is an example of optimization of the weights (w l ) l=1:::L . A similar "softmax" combination of local... |

41 |
Discrete Parameter Variation: Efficient Estimation of a Switching Regression Model
- Kiefer
- 1978
(Show Context)
Citation Context ...ch Gaussian modeling a subregion and having distribution parameters initialized from the clustering procedure. The idea underlying the model is strongly related to the concept of switching regression =-=[7, 8] and to the "mi-=-xture of expert" approach [9, 10]. The Expectation step in the EM algorithm [7, 9] is an example of optimization of the weights (w l ) l=1:::L . A similar "softmax" combination of local... |

40 | A baseline of a speaker independent continuous speech recognizer of italian
- Angelini, Brugnara, et al.
- 1993
(Show Context)
Citation Context ...-dependent, HMMbased) to normalize the input acoustic data from a new speaker. The experiments were performed on the speaker-dependent portion of the acoustic-phonetic continuous speech corpus APASCI =-=[6]-=-. The mapping was realized with a single, global GRAN network trained on 40 sentences: phone recognition error was reduced with an average adaptability ratio of 25% (see later for a definition of this... |

19 |
Speaker normalization for speech recognition
- Huang
- 1992
(Show Context)
Citation Context ...ence speaker (for whom the speech recognition system is trained) is also called spectral mapping [1, 2]. A similar mapping has been synthesized by Multi-Layer Perceptron networks by other researchers =-=[3, 4]-=-. A previous paper introduced GRAN networks for speaker normalization [5]. The neural network was applied as a front-end preprocessor for a continuous speech recognition system (speaker-dependent, HMM... |

12 |
Speaker normalization and adaptation using second-order connectionist networks
- Watrous
- 1993
(Show Context)
Citation Context ...ence speaker (for whom the speech recognition system is trained) is also called spectral mapping [1, 2]. A similar mapping has been synthesized by Multi-Layer Perceptron networks by other researchers =-=[3, 4]-=-. A previous paper introduced GRAN networks for speaker normalization [5]. The neural network was applied as a front-end preprocessor for a continuous speech recognition system (speaker-dependent, HMM... |

10 |
Mixture models, outliers, and the EM algorithm
- Aitkin, Tunnicliffe-Wilson
- 1980
(Show Context)
Citation Context ...tribution parameters initialized from the clustering procedure. The idea underlying the model is strongly related to the concept of switching regression [7, 8] and to the "mixture of expert"=-= approach [9, 10]. The Expe-=-ctation step in the EM algorithm [7, 9] is an example of optimization of the weights (w l ) l=1:::L . A similar "softmax" combination of local linear function was realized in [11], with weig... |

7 |
Fast speaker adaptation for speech recognition
- Class, Kaltenmeier, et al.
- 1990
(Show Context)
Citation Context ...del. The mapping function between the acoustic parameter space of the new speaker and that of the reference speaker (for whom the speech recognition system is trained) is also called spectral mapping =-=[1, 2]-=-. A similar mapping has been synthesized by Multi-Layer Perceptron networks by other researchers [3, 4]. A previous paper introduced GRAN networks for speaker normalization [5]. The neural network was... |

7 |
Robust speaker adaptation using a piecewise linear acoustic mapping
- Bellegarda, Souza, et al.
- 1992
(Show Context)
Citation Context ...del. The mapping function between the acoustic parameter space of the new speaker and that of the reference speaker (for whom the speech recognition system is trained) is also called spectral mapping =-=[1, 2]-=-. A similar mapping has been synthesized by Multi-Layer Perceptron networks by other researchers [3, 4]. A previous paper introduced GRAN networks for speaker normalization [5]. The neural network was... |

7 | Connectionist speaker normalization with generalized resource allocating networks
- Furlanello, Giuliani, et al.
- 1995
(Show Context)
Citation Context ...led spectral mapping [1, 2]. A similar mapping has been synthesized by Multi-Layer Perceptron networks by other researchers [3, 4]. A previous paper introduced GRAN networks for speaker normalization =-=[5]-=-. The neural network was applied as a front-end preprocessor for a continuous speech recognition system (speaker-dependent, HMMbased) to normalize the input acoustic data from a new speaker. The exper... |

2 |
Voice adaptation using multi-functional transformation with weighting by radial basis function networks
- Iwahashi, Sagisaka
- 1994
(Show Context)
Citation Context ...pproach [9, 10]. The Expectation step in the EM algorithm [7, 9] is an example of optimization of the weights (w l ) l=1:::L . A similar "softmax" combination of local linear function was re=-=alized in [11], with wei-=-ghts estimated by an RBF network, while reference [12] gives indications for a probabilistic treatment of the general case. Here the joint use of this "architectural" approach is exploited c... |

2 |
Applied nonparametric regression, Vol. 19 of Econometric Society Monograph
- Härdle
- 1990
(Show Context)
Citation Context ...rates for each type of parameter. Implemented kernels are Gauss, Hardy multiquadrics, inverse of Hardy multiquadrics and Epanenchnikov kernels, also in the Nadaraya-Watson ("softmax") normal=-=ized form [16]-=-. Classical RBF learning, i.e. kernel placement via clustering and computation of kernel coefficients via SVD estimation is also available. Specific architectural, parameter and learning choices for t... |