home   people   research   events   teaching   vacancies   links

Assessing the reverberation level in speech

Problem
In most real-world situations a sound signal is affected by room acoustics, that is, reflections of the original signal distort the total sound signal. While humans can cope with the effects of reverberation up to a high level, the performance of artificial systems, such as automatic speech recognition (ASR) systems, is seriously degraded even by low levels of reverberation.

Possible solution
Several methods have been developed to cope with the effects of reverberation. For example, speech enhancement methods try to recover the direct speech from the total signal. Another possibility is to train the ASR system on reverberant speech. However, the effects of room acoustics vary greatly for different environments. When the room parameters (size, absorbtion, etc.) are known, and are constant, there is no problem, but most real-world situations do not fulfill this requirement. A good estimation of the reverberation level (usually expressed by the reverberation time, T60, the time it takes for the total energy to decay 60 dB) is then needed for these methods to work. Standard methods for the estimation of reverberation time need active measurements, so they are neither blind nor real-time.

Proposal
We would like to develop and test features for online and blind estimation of reverberation time. Some work has already been done (see this paper, and this poster) and will serve as a starting point. An important part of the research will be the development and execution of experiments for correct measurement of reverberation times. Furthermore, the existing features may be improved and extended. Of course, any new ideas are welcome.

Interested?
Mail or visit Maria (, BB room 326)