Master 2013 2014
Stages de la spécialité SAR
Development of procedures for efficient retraining of acoustic recognizers

Site :Fraunhofer IDMT, Hearing Speech and Audio Technology
Lieu :Oldenburg, Germany
Encadrant : Danilo Hollosi, Jens Schröder,
Dates :To be determined.
Rémunération :500€ Monthly
Mots-clés : Parcours ATIAM : Traitement du signal


At the Fraunhofer Institute For Digital Media Technology, Project Group Hearing, Speech and Audio Technology, machine learning approaches are used for recognizing speech and acoustic events e.g. for machine and facility monitoring. Therefore, an acoustic utterance recorded by microphones is transcribed into features representing the acoustic class or speech command. These features are fed to a machine learning method that trains models to automatically classify acoustic events. Most often, this training is very time consuming and to consider new data in a model, the whole training step for this model has to be repeated. The goal for this thesis is to develop a method that is capable of efficiently retraining Hidden Markov Models (HMMs) and a prototypic implementation in MATLAB. Together with the Fraunhofer Project Group, an application area will be defined beforehand. The new approach has to be evaluated and compared to existing methods in consideration of accuracy and speed.