Probabilistic analysis of pronunciation with MAUS

This paper describes a method to auto-matically detect pronunciation variants in large speech corpora within the frame-work of the 'MAUS ' project ([1]). 'MAUS' stands for 'Munich Automatic Segment-ation System ' and is a general purpose tool to automatically label and segment read and spontaneous German speech into phonetic/phonologic segments. The out-put of MAUS can for example be used to build probabilistic models of pronun-ciation of uent German as re ected by the analysed corpus. These models can be the basis for phonetic investigations or can be incorporated into classic speech re-cognition algorithms. The paper is organised as follows: The rst section gives a very short introduc-tion into the main processing principle of MAUS and gives some examples of the output of MAUS applied to utter-ances from the Verbmobil corpus. Sec-tion 2 deals very brie y with the prob-lem of how to evaluate such an output. A method is given that rst compares the performance of three human transcribers with each other and then the performance of MAUS with each of them. Section 3 describes our method for deriving prob-abilistic pronunciation dictionaries from the MAUS output and gives some inter-esting examples from the Verbmobil do-main. The 4th and last section gives some new approaches towards incorpor-ating these models into a new automatic speech recognition (ASR) approach that combines phonetically 'sharper ' acoustic models with the probabilistic modelling of pronunciation