Intonation: Theory, Models, and Applications

Athens, Greece
September 18-20, 1997


Some Experiments About the Use of Prosodic Parameters in a Speech Recognition System

Katarina Bartkova

France Telecom CNET, DIH/RCP, Technopole Anticipa, Lannion Cedex, France

This paper summarises results of experiments related to the introduction of prosodic parameters in automatic speech recognition in French. The goal of using prosodic parameters was to improve speech recognition in real applications such as vocal services available in the telephone network [1]. The prosodic parameters studied were phoneme duration, pitch movement and phoneme intensity. They were used in a post-processing procedure of an automatic speech recognition system based on HMM modelling. The first set of experiments involved re-scoring the N-best solutions, supplied by the HMM modelling in order to reduce the recognition error rate. The second set of experiments aimed to improve, in a real application context, the rejection of incorrect tokens such as noise tokens and out-of-vocabulary words. It was shown that the prosodic parameters can be efficiently used to improve recognition performance either by recovering the HMM errors or by improving the rejection of incorrect tokens.

Full Paper

Bibliographic reference.  Bartkova, Katarina (1997): "Some experiments about the use of prosodic parameters in a speech recognition system", In INT-1997, 33-36.