Neighborhood effects on spoken word recognition in Japanese

Abstract

In speaker verification, two independent stochastic models, i.e. a client model and a non-client (world) model, are generally used to verify the claimed identity using a likelihood ratio score. This paper investigates a variant of this approach based on a common hidden process for both models. In this framework, both models share the same topology, which is conditioned by the underlying phonetic structure of the utterance. Then, two different output distributions are defined corresponding to the client vs. world hypotheses. Based on this idea, a synchronous decoding algorithm and the corresponding training algorithm are derived. Our first experiments on the SESP telephone database indicate a slight improvement with respect to a baseline system using independent alignments. Moreover, synchronous alignment offers a reduced complexity during the decoding process. Interesting perspectives can be expected.

Farsi language prosodic structure, research and implementation using a speech synthesizer

“…A first set of recent experimental results (also reported in [5]) show a slight benefit in terms of EER for the synchronous alignment approach. Moreover, the approach offers a slightly reduced computational complexity and provides a simple decomposition of the utterance log likelihood ratio in terms of a sum of frame-by-frame likelihood ratios.…”

“…The use of synchronous decoding turns out to be more consistent if similar constraints have been introduced during the training of the client model. Detailed decoding and training procedures for the synchronous alignment scheme can be found in [5].…”