Interspeech'2005 - Eurospeech
The aim of the present investigation was to find characteristics in the speech signal that allow distinguishing the German obstruents /pf/ and /ts/, /f/ and /s/ and /p/ and /t/. Two acoustic analysis techniques were chosen to separate the target phonemes according to place and manner of articulation: measuring duration distinguished them in manner (fricative, stop and affricate), whereas the calculation of the relative amplitude of particular frequency bands distinguished them according to place of articulation (labial and alveolar). The speech data was recorded in a natural dialog situation to gain features that are robust towards variation in the speech signal. The characteristics found will serve for the improvement of the "FUL" automatic speech recognition system . The FUL speech recognition system is based on underlying phonological features providing robustness for all kinds of variations in the speech signal.
Bibliographic reference. Hoelterhoff, Julia (2005): "Recognition of German obstruents", In INTERSPEECH-2005, 2917-2920.