Click for new scientific resources and news about Corona[COVID-19]

Paper Information

Title: 

SPEAKER VARIABILITY COMPENSATION BY INPUT ADAPTATION

Type: PAPER
Author(s): NEZHADGHOLI ISAR,SEYYED SALEHI SEYYED ALI
 
 
 
Name of Seminar: IRANIAN CONFERENCE ON BIOMEDICAL ENGINEERING
Type of Seminar:  CONFERENCE
Sponsor:  ANJOMANE MOHANDESI PEZESHKI IRAN
Date:  2004Volume 11
 
 
Abstract: 

ACOUSTIC VARIABILITY ACROSS SPEAKER IS ONE OF THE CHALLENGES OF SPEAKER INDEPENDENT SPEECH RECOGNITION SYSTEMS. IN THIS PAPER AN INPUT ADAPTATION METHOD IS PROPOSED TO COMPENSATE THE EFFECT OF THIS VARIABILITY. THE MODEL USED HERE IS A TIME DELAY NEURAL NETWORK (TDNN). THE TDNN IS TRAINED BASED ON SPEAKERS IN TRAIN SET. AMONG THESE 71 SPEAKERS, THE BEST ONE (WITH HIGHEST PHONE RECOGNITION ACCURACY) IS SELECTED AS "REFERENCE SPEAKER". THEN USING BACK PROPAGATION ALGORITHM, ALL SPEECH SIGNALS IN TRAIN SET ARE ADAPTED TO FIT THIS SPEAKER'S SPEECH SIGNAL AND A NEURAL NETWORK IS TRAINED BASED ON ADAPTED DATA. TO EVALUATE THE PERFORMANCE OF THE TRAINED NETWORK, TEST SIGNALS RELATED TO 30 SPEAKERS ARE ADAPTED TO REFERENCE SPEAKER, EXACTLY THE SAME AS TRAIN SPEAKERS. BY IMPLEMENTING THIS METHOD AND COMBINING THE RESULTS OF ADAPTED AND UNADAPTED NETWORKS REGARDING MAXIMUM CONFIDENCE LEVEL, 2.6% IMPROVEMENT IS GAINED IN PHONE RECOGNITION ACCURACY.

 
Keyword(s): 
 
 
Yearly Visit 43   tarjomyar
 
Latest on Blog
Enter SID Blog