Effect of frequency weighting on MLP-based speaker canonicalization

Accurate and efficient speaker canonicalization is proposed to improve the performance of speaker-independent ASR systems. Vocal tract length normalization (VTLN) is often applied to speaker canonicalization in ASR; however, it requires parallel decoding of speech when estimating the optimal warping parameter. In addition, VTLN provides the same linear spectral transformation in an utterance, although optimal mapping functions differ among phonemes. In this study, we propose a novel speaker canonicalization using multilayer perceptron (MLP) that is trained with a data set of vowels to map an input spectrum to the output spectrum of a standard speaker or a canonical speaker. The proposed speaker canonicalization operates according to the integration of MLP-based mapping and identity mapping that depends on frequency bands and achieves accurate recognition without any tuning of mapping function during run-time. Results of experiments conducted with a continuous digit recognition task showed that the proposed method reduces the intra-class variability in both of the vowel and consonant parts and outperforms VTLN.