Back to search

Technical University of Crete (TUC)

Maximum-likelihood stochastic transformations for continuous speech recognition

Abstract

dc:description

This thesis presents our efforts to address two major problems in current large vocabulary continuous speech recognition systems. The first problem is the mismatched conditions between the training and testing sets. We particularly focus on the performance degradation due to different speakers and dialects. The second problem is the explicit modeling of the inter-frame correlations in a speaker-independent (SI) system. We attack both of these problems by applying strategies based on the popular family of linear model transformations and we further propose a novel stochastic transformation scheme named Maximum Likelihood Stochastic Transformations (MLST).MLST consists of a collection of component, linear transformations that are shared among all the Gaussians in each mixture. The component transformation applied to each Gaussian is selected probabilistically, based on weight probabilities. For the estimation of the transformation parameters and weight probabilities we use the expectation-maximization (EM) algorithm. The single linear transformations can then be seen as a special case of MLST with one linear component. In its generic form, MLST is a more advantageous transformation since it provides the mechanism to increase acoustic resolution while maintaining the robust estimation of its parameters. Although the component transformations may be tied over a set of mixtures, the choice of the specific transformation that is applied to a Gaussian is based on a soft decision process using the transformation weights.To evaluate our new method we first perform a series of speaker-adaptation experiments using SRI's DECIPHER speech recognition system. We find that our new method significantly outperforms the methods based on the single linear transformations. Furthermore, the overall computational demands remain almost the same as with the linear approach. We then investigate the effect of mismatched dialects and find that it significantly influences the recognition accuracy. Consequently, we apply several adaptation approaches, which had been previously used in the speaker or channel adaptation tasks, to develop a dialect-specific recognition system using a dialect-dependent system trained on a different dialect and a small number of training sentences from the target dialect. We show that adaptation improves recognition performance dramatically and increases robustness with small amounts of training sentences. Furthermore, using MLST in dialect-adaptation experiments we get significant performance improvement over the linear approaches.Finally, we incorporate the MLST transformations in the training process of SI systems in an effort to explicitly model the correlations between the spectral parameters. We find that, although some performance gains can be obtained, performance is not significantly influenced by the number of MLST component transforms, nor by the type of the transformation.

Degree

thesis:*
Grantor dc:publisher
Technical University of Crete (TUC)
Year dc:date
2000

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Diakoloukas, Vasileios
  • Διακολουκάς, Βασίλειος

Subjects

dc:subject × 16

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/32036

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Diakoloukas, Vasileios; Διακολουκάς, Βασίλειος. Maximum-likelihood stochastic transformations for continuous speech recognition. Technical University of Crete (TUC), 2000. http://hdl.handle.net/10442/hedi/32036