Multi-speaker/Speaker-independent Architectures for the Multi-state Time Delay Neural Network

Abstract

In this paper we present an improved Multi-State Time Delay Neural Network (MS-TDNN) for speaker-independent, connected letter recognition which outperforms an HMM based system (SPHINX) and previous MS-TDNNs, and explore new network architectures with "internal speaker models". Four different architectures characterized by an increasing number of speaker-specific parameters are introduced. The speaker-specific parameters can be adjusted by "automatic speaker identification" or by speaker adaptation, allowing for "tuning-in" to a new speaker. Both methods lead to significant improvements over the straightforward speaker-independent architecture. Similar as described in [BRIDLE91], even unsupervised "tuning-in" (speech is unlabeled) works astonishingly well.