The segregation of spatialised speech in interference by optimal mapping of diverse cues

TY - GEN

T1 - The segregation of spatialised speech in interference by optimal mapping of diverse cues

AU - Gao, Jingbo

AU - Tew, Anthony I

PY - 2015/4/19

Y1 - 2015/4/19

N2 - We describe optimal cue mapping (OCM), a potentially real-time binaural signal processing method for segregating a sound source in the presence of multiple interfering 3D sound sources. Spatial cues are extracted from a multisource binaural mixture and used to train artificial neural networks (ANNs) to estimate the spectral energy fraction of a wanted speech source in the mixture. Once trained, the ANN outputs form a spectral ratio mask which is applied frame-by-frame to the mixture to approximate the magnitude spectrum of the wanted speech. The speech intelligibility performance of the OCM algorithm for anechoic sound sources is evaluated on previously unseen speech mixtures using the STOI automated measures, and compared with an established reference method. The optimized integration of multiple cues offers clear performance benefits and the ability to quantify the relative importance of each cue will facilitate computationally efficient implementations.

AB - We describe optimal cue mapping (OCM), a potentially real-time binaural signal processing method for segregating a sound source in the presence of multiple interfering 3D sound sources. Spatial cues are extracted from a multisource binaural mixture and used to train artificial neural networks (ANNs) to estimate the spectral energy fraction of a wanted speech source in the mixture. Once trained, the ANN outputs form a spectral ratio mask which is applied frame-by-frame to the mixture to approximate the magnitude spectrum of the wanted speech. The speech intelligibility performance of the OCM algorithm for anechoic sound sources is evaluated on previously unseen speech mixtures using the STOI automated measures, and compared with an established reference method. The optimized integration of multiple cues offers clear performance benefits and the ability to quantify the relative importance of each cue will facilitate computationally efficient implementations.

KW - Speech segregation

KW - Artificial Neural Networks

KW - ratio mask

U2 - 10.1109/ICASSP.2015.7178340

DO - 10.1109/ICASSP.2015.7178340

M3 - Conference contribution

SN - 978-1-4673-6997-8

SP - 2095

EP - 2099

BT - 2015 IEEE International Conference on Acoustics, Speech, and Signal Processing

PB - IEEE

CY - Brisbane

ER -