Grants and Contracts Details
Description
The proposed work focuses on developing masking algorithms to extract weak voices in cocktail
party environments using distributed microphones. This builds on previous work where it was
demonstrated that signals from distributed microphones can be used to create time-frequency
masking patterns that suppress interfering voices and significantly increase intelligibility over
beamforming alone. The student funded by this proposal will develop algorithms to reduce the
need for knowing precise microphone locations through auto-tuning the steering vectors, and
will explore other time-frequency decomposition methods with the potential to reduce distortion
from the masking operations. While previous/current work has considered auto-tuning, more
refinement in the optimization processes is required to find the weak local minima in the steeredresponse
power of the distributed microphone beamformer, which represents the weaker targets.
In addition, current time-frequency masking uses regular time-frequency intervals based on an
short-time FFT decomposition. Wavelet packets have the potential window the time-frequency
space in a way more compatible with speech signals and human auditory processing. The use of
wavelet packet will be explored to determine best basis patterns for improving intelligibility and
reducing distortion. Experiments based on real recordings and simulations will be performed for
developing and testing performance. Best performing algorithms will be packaged in graphical
user interfaces for use by the sponsor.
Status | Finished |
---|---|
Effective start/end date | 6/4/12 → 7/3/13 |
Fingerprint
Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.