Unsupervised Cross-Adaptation Approach for Speech Recognition by Combined Language Model and Acoustic Model Adaptation

The aim of this study is to improve speech recognition with a combination of language model (LM) and the acoustic model (AM) adaptation. The proposed adaptation techniques are based on cross-system adaptation or cross-validation (CV) adaptation. The principle is to use complementary information derived from several systems or data sets. Because language information and acoustic information differ completely, the combined approach is expected to be effective. We evaluate the performance of the proposed methods by conducting speech recognition experiments using the Corpus of Spontaneous Japanese (CSJ). Both cross-system adaptation and CV adaptation give better performance than the conventional adaptation method; the crosssystem adaptation method was found to exhibit the best recognition performance.