Abstract:

Modern machine learning algorithms demand large amounts of compuational resources. The Vienna Scientific Cluster (VSC) provides such resources to researchers. This thesis should serve as an inital project to transfer automatic speech recognition (ASR) experiments to the VSC. This thesis is intended as a group project for 2-3 students.

Your Tasks:

• get to know the VSC environment (recommendation: attend the online course ‘‘Introduction to Working on the VSC Clusters’’)
• setup/configure example tasks with the speech processing frameworks fairseq and Kaldi
• reproduce baseline speech recognition results for conversational Austrian German (GRASS) with both systems (scripts available) with focus on distributed training
• setup/train transformer language models (LM) with GRASS by finetuning pre-trained models
• compare existing decoding strategies of the available ASR pipeline with your LM implementations
• provide an extensive documentation of your experiences setting up speech recognition experiments on the VSC

Your Profile

• good knowledge in programming (Shell/Python/…) and UNIX tools
• good knowlegde in data science and machine learning tooling

Contact:

Julian Linke (linke@tugraz.at) Philipp Gabler (pgabler@know-center.at)