Nikko STROM
A Speaker Sensitive
Artificial Neural Network Architechture
for Speaker Adaptation
Abstract:The speaker sensitive framework for speaker adaptation is outlined and exemplified by a speaker sensitive artificial neural network. A technique for automatically
extracting the speaker-characteristics space is introduced and evaluated on a vowel
discrimination task. The generated speaker-space is analyzed using analysis-by-
synthesis and the shifts found in the F1/F2-space are consistent with results from
speech production research. The speaker-space is also compared with, and found to
be correlated with, the knowledge-based speaker parameters F0 and spectral tilt.