Introducing Block-Wise Processing into Turbo Viterbi ASR

Inhalt:
Recently, turbo automatic speech recognition (ASR) has been proposed as intermediate-level fusion approach for multi-channel, or multi-model, or multi-modal ASR. To apply turbo ASR in a continuous recognition application, the turbo Viterbi ASR has to be used. Working towards turbo ASR for real-time implementations in a practical context, the contributions of this paper are the following: After briefly reviewing the turbo Viterbi ASR approach, we extend it by introducing a new block-wise processing of the iterations, and a new simplified weighting scheme for both the likelihoods and the so-called extrinsic information. Applied to a single channel unimodal ASR task with fusion of two types of features, we analyze and discuss the proposed extensions with respect to latency and required complexity.