Room Classification from Reverberant Speech using Relative Transfer Function
Classifying rooms based on their volume or geometry can be useful for scene analysis
and characterization of the recording’s environment. Potential applications may
include forensics, robot navigation, and location-aware voice interfaces.
In this thesis, we present a room classification algorithm using speech signals.
We classify rooms which differentiate by volume and geometry using speech signals
recorded in the room. The Relative Transfer Function (RTF) is proposed as a new
feature vector which exploits the information contained in an array of microphones,
unlike other methods, which either use only one microphone, or is based on the
Room Impulse Response (RIR), which is difficult to estimate. We also investigate
the use of Neural Networks (NNs) as a classifier and their ability to learn a better
representation of the room.
To evaluate our method we simulated data using a room impulse response generator and compared our results to other methods under the same conditions in a
variety of scenarios. We have found that our method outperforms the competing
methods in some of the scenarios.
* M.Sc. research supervised by Prof. Sharon Gannot