Segment-based speaker adaptation by neural network

Abstract

The authors propose a segment-to-segment speaker adaptation technique using a feed-forward neural network with a time shifted sub-connection architecture. Differences in voice individuality exist in both the spectral and temporal domains. It is generally known that frame based speaker adaptation techniques can not compensate for speaker individuality in the temporal domain. Segment based speaker adaptation compensates for these spectral and temporal differences. The results of 23 Japanese phoneme recognition experiments using TDNN (time-delay neural network) show that the recognition rate by segment-based adaptations was 83.7%, 22.8% higher than the rate without adaptation.<>