Comparison of predictive measures of speech recognition after noise reduction processing
2014 (English)In: Journal of the Acoustical Society of America, ISSN 0001-4966, E-ISSN 1520-8524, Vol. 136, no 3, 1363-1374 p.Article in journal (Refereed) Published
A number of measures were evaluated with regard to their ability to predict the speech-recognition benefit of single-channel noise reduction (NR) processing. Three NR algorithms and a reference condition were used in the evaluation. Twenty listeners with impaired hearing and ten listeners with normal hearing participated in a blinded laboratory study. An adaptive speech test was used. The speech test produces results in terms of signal-to-noise ratios that correspond to equal speech recognition performance (in this case 80% correct) with and without the NR algorithms. This facilitates a direct comparison between predicted and experimentally measured effects of noise reduction algorithms on speech recognition. The experimental results were used to evaluate nine different predictive measures, one in two variants. The best predictions were found with the Coherence Speech Intelligibility Index (CSII) [Kates and Arehart (2005), J. Acoust. Soc. Am. 117(4), 2224-2237]. In general, measures using correlation between the clean speech and the processed noisy speech, as well as other measures that are based on short-time analysis of speech and noise, seemed most promising.
Place, publisher, year, edition, pages
2014. Vol. 136, no 3, 1363-1374 p.
Speech recognition, Noise reduction
Other Electrical Engineering, Electronic Engineering, Information Engineering
Research subject Speech and Music Communication
IdentifiersURN: urn:nbn:se:kth:diva-149647DOI: 10.1121/1.4892766ISI: 000342205700047PubMedID: 25190409OAI: oai:DiVA.org:kth-149647DiVA: diva2:750297
QC 201410212014-09-282014-08-252014-10-21Bibliographically approved