Mean and variance adaptation within the MLLR framework

摘要: Abstract One of the key issues for adaptation algorithms is to modify a large number parameters with only small amount data. Speaker techniques try obtain near speaker-dependent (SD) performance amounts speaker-specific data, and are often based on initial speaker-independent (SI) recognition systems. Some these speaker may also be applied task new acoustic environment. In this case an SI system trained in, typically, clean environment adapted operate in new, noise-corrupted, This paper examines maximum likelihood linear regression (MLLR) technique. MLLR estimates transformations groups model maximize Previously, has been mean mixture-Gaussian HMM extended update Gaussian variances re-estimation formulae derived variance transforms. compensation evaluated several vocabulary tasks. The use was found give additional 2% 7% decrease word error rate over mean-only adaptation.