Design of multi-feature class models for Speech Recognition Security systems with under-resourced languages

摘要

One of the goals of Speech Recognition Security (SRS) systems is to have appropriately tools to recognize speech password spoken based on elements such as words, sub-word or speakers. The main goal of the present work is to design robust ASR systems based on alternative ways to the classical evaluation rates, which often depend on the vocabulary of the task and on the language resources available. The drawback of this approach is that it is not straightforward that a system with a slightly lower WER during tests will adapt properly to new utterances, and this is much more sensible when the baseline system has a big error rate since there are many features that could be improved. This tends to be the case of under-resourced languages, since the lack of resources has a great impact in the performance of the system and not all the standard methods are suitable to any kind of language or task. The novel approach is to choose balanced multi-features of the acoustic models and the sub-word units based on rates related to entropy, mutual information and similitude. Selected models are integrated in an ontology-driven Audio Information Retrieval system that suits the requirements of under-resourced languages.