Konferenzbeitrag
Automatic parameter tuning and extended training material: recent advances in the Fraunhofer speech recognition system
Vorschaubild nicht verfügbar
Volltext URI
Dokumententyp
Text/Conference Paper
Dateien
Zusatzinformation
Datum
2013
Autor:innen
Zeitschriftentitel
ISSN der Zeitschrift
Bandtitel
Verlag
Gesellschaft für Informatik e.V.
Zusammenfassung
Building the acoustic and language models on a larger amount of training data is a well-known method for robustifying automatic speech recognition approaches. The adaption of the decoder settings afterwards, however, is often only marginally addressed (e.g. being manually set or using default values provided by a toolkit). Without proper adaption, these settings are most often sub-optimal and lead to degraded performance without unlocking the full potential of the speech recognizer. Ideally, the decoder settings should be optimized after each modification of the language model and/or the acoustic model of the speech recognition system, a task that is typically too tedious for manual work. In this paper, we employ an automatic optimization technique on the Fraunhofer IAIS speech recognition setup as a subsequent step to training data increase. We will present the improvements of the expanded training data for the acoustic models and the optimization of the decoder settings on the German Difficult Speech Corpus. Index Terms: simultaneous perturbation stochastic approximation, free decoding parameters, word error rate optimization