Lombard speech recognition based on voiced sound detection and application to the fabric inspection system in factories
Sukeyasu Kanno, T. Funada · Systems and Computers in Japan · 2003
Abstract In a very noisy environment such as a factory, the performance of speech recognition drops due to the presence of noise and the Lombard effect. Given this, the present authors consider a method to recognize noisy Lombard speech as based on the effective detection of voiced sound period under noisy conditions in a factory. Based on recognition experiments using a 120‐word vocabulary spoken by three men, the authors showed that recognition performance for noisy Lombard speech was high when the WGD measure used speech under noisy conditions as a reference pattern. Then the authors confirmed, based on the proposals in this paper, the effects of weighting distance in the voiced sound period and the effectiveness of using word periods based on voiced sound. In addition, the authors developed an inspection system using speaker‐dependent speech recognition based on the results of these evaluations. The results of operational experiments to input the names of fabric defects (a 61‐word vocabulary) by voice in an inspection factory (noise level: 75 to 90 dBA) yielded a recognition rate of 93.1%, clearly improving operating efficiency. © 2003 Wiley Periodicals, Inc. Syst Comp Jpn, 34(7): 10–23, 2003; Published online in Wiley InterScience ( www.interscience.wiley.com ). DOI 10.1002/scj.10369