Motivated by two applications from the field of music information retrieval 
(MIR), audio identification and audio matching,
we address some general principles for constructing robust and, 
at the same time, semantically meaningful audio features. 
Whereas audio identification aims at identifying a short excerpt
(let's say of about 10-30 seconds of duration) of
audio as being part of a particular audio recording taken
from a particular CD, audio matching aims at 
automatically retrieving all musically similar excerpts in all
interpretations of the underlying pieces of music, which are 
contained in the database. Audio matching may hence be considered
as a sematically advanced retrieval problem.
In our talk, we first consider an approach to feature design
based on calculating time-domain statistics for the well-known class of
pitch-based chroma features. This allows us to successfully approach
the problem of audio matching for the case of western classical
music. Subsequently, we adapt the chroma-approach to extract
robust tempo-related features resulting in the concept of a
cyclic beat spectrum (CBS). Using a suitable feature set derived
from the CBS, we propose an approach to robust time-scale invariant
audio identification.