Acoustic and language model adaptation in a voice interactive system for elderly people - Archive ouverte HAL Access content directly
Conference Papers Year : 2015

Acoustic and language model adaptation in a voice interactive system for elderly people

(1) , (2, 3, 4, 5) , (6, 7) , (6, 7) , (1)
1
2
3
4
5
6
7

Abstract

Automatic Speech Recognition (ASR) systems can perform better if trained for a specific application. Though, since we require a huge amount of information to train models it is not feasible to build such systems once ready for the user, but we could use adaptation to make the ASR system more appropriate for the final use. In this work we address adaptation for the vocal characteristics of the speaker, environmental noise and the language model. Acoustic model adaptation is done by Speaker Adaptive Training (SAT), linear Vocal Tract Length Normalization (lVTLN) and constrained Maximum Likelihood Linear Regression (cMLLR). Interpolation is applied for language model adaptation. The relative WER reduction by using cMLLR was9.44%. The perplexity of the language model could be relatively improved by 14.47%
Fichier principal
Vignette du fichier
JETSAN-2015_Saeideh.pdf (156.16 Ko) Télécharger le fichier
Origin : Files produced by the author(s)
Loading...

Dates and versions

hal-01273059 , version 1 (11-02-2016)

Identifiers

  • HAL Id : hal-01273059 , version 1

Cite

Saeideh Mirzaei, Jérôme Boudy, Pierrick Milhorat, Gérard Chollet, Mikko Kurimo. Acoustic and language model adaptation in a voice interactive system for elderly people. JETSAN 2015 : 5ème colloque des Journées d'Etudes sur la TéléSanté, May 2015, Compiègne, France. pp.1 - 4. ⟨hal-01273059⟩
284 View
79 Download

Share

Gmail Facebook Twitter LinkedIn More