Modèles de production de parole et reconnaissance à partir d'automates

Abstract : The framework of this thesis is speaker-independent automatic speech recognition. This document begins with a short survey of speech signal processing applied to speech recognition. Then, we describe several common architectures: dynamic time warping of acoustic shapes, artificial intelligence, neural networks and hidden Markov models. This document is made of two main parts. The first part is devoted to the recognition of words. We are using finite state automata for modeling the eleven American spoken digits. As the speech database TiDigits has been used by other teams we can compare our results against thoose obtained with other approaches. The first step is concerned with isolated word recognition. Then, we describe how sentences of the database have been segmented. Last sections of this part describe continuous speech recognition of word sequences, as well as a discussion about strong and weak points of our approach. The second part treats of the use of production models for speech recognition. We begin with a description of the acoustic equations which drive the air flow inside the vocal tract and we present several articulatory models. Then, we justify the choice of Maeda's model. We describe the adaptation of this model to a male speaker. Next, we describe the variational method used for recovering articulatory trajectories from the speech. Finally, the software we built, is described. In the conclusion, we give the results obtained and we exhibit sorne perspectives for future work towards a better speaker indepedent continuous speech recognition system.
Document type :
Theses
File URL :
http://docnum.univ-lorraine.fr/prive/SCD_T_1999_0300_MATHIEU.pdf
Complete list of metadatas

https://hal.univ-lorraine.fr/tel-01747264
Contributor : Thèses Ul <>
Submitted on : Thursday, March 29, 2018 - 10:56:00 AM
Last modification on : Tuesday, April 24, 2018 - 1:30:16 PM

Identifiers

  • HAL Id : tel-01747264, version 1

Collections

Citation

Bruno-Charles Mathieu. Modèles de production de parole et reconnaissance à partir d'automates. Autre [cs.OH]. Université Henri Poincaré - Nancy 1, 1999. Français. ⟨NNT : 1999NAN10300⟩. ⟨tel-01747264⟩

Share

Metrics

Record views

12