5th International Conference on Spoken Language Processing
We present an approach to joint application of spectral subtraction (SPS) and model combination (PMC) for speech recognition in noisy environments. Contrary to previous solutions distortion introduced by SPS is not modeled in PMC. Instead we ensure compatibility of the two methods by adapting parameters of SPS (spectral floor and overestimation factor) according to the present signal-to-noise-ratio (SNR). Parameter setting should be done to subtract a maximum of noise while minimizing distortion. Experiments suggest that for each noise level different parameter sets yield optimal performance. Setting the parameters adaptively according to the noise level leads to undegraded results at high SNR while in low SNR regions the benefits of the noise reduction process are significant. This scheme leaves the model combination process unchanged which simplifies parameter estimation and reduces computation time. Experiments show significant improvements when using PMC with modified SPS instead of standard SPS.
Bibliographic reference. Schless, Volker / Class, Fritz (1998): "SNR-dependent flooring and noise overestimation for joint application of spectral subtraction and model combination", In ICSLP-1998, paper 0138.