Sixth European Conference on Speech Communication and Technology
Real-world speech and speaker recognition systems are often subject to ambient noise which results in significant performance loss, more so when the noise types and noise levels are different between training and testing. This paper presents a new preprocessing technique, Coherent Spectral Modification, that aims to reduce distortion due to noise by modifying the complex speech spectrum using information from non-speech regions of the spectral time trajectories. A refinement process is also proposed that further reduces noise mismatch. This combined technique was evaluated on a speaker verification task where the test data was corrupted with varying levels of white noise and pink noise. The new method yielded significant reduction in error rates and performed better than conventional spectral subtraction, particularly at moderate SNRs.
Full Paper (PDF) Gnu-Zipped Postscript
Bibliographic reference. Ramanujam, Vidhya / Balchandran, Rajesh / Mammone, Richard J. (1999): "Robust speaker verification in noisy conditions by modification of spectral time trajectories", In EUROSPEECH'99, 791-794.