RESPITE: The CASA Toolkit Page: Documentation: Block Library Index:GMask


The GMask block produces `cleaned' speech, the present data mask and the probability of the present mask, using the noisy speech and the noise means and variance estimates. This block is intended to be used together with blocks such as the AdaptiveNoiseEstimation block which produce estimates of the noise mean and variance. It is assumed that the noisy data is spectral data, that the distribution of the estimated noise is Normal, and that the channels are independent.

The `cleaned' speech (out1) is simply the noisy speech minus the noise mean estimate. The probability of present data (out3) is P(estimated_SNR<NOISE_THRESHOLD), and the present data mask (out2) is computed as P(estimated_SNR<NOISE_THRESHOLD) > PROBABILITY_THRESHOLD.

WARNING: the cleaned speech may contain negative values!"

Inputs Meaning Sample 1-D frame $\ge$2-D frame
in1 noisy data Yes Yes Yes
in2 estimate of noise mean Yes Yes Yes
in3 estimate of noise variance Yes Yes Yes

Outputs Meaning
out1 `cleaned' speech
out2 present data mask
out3 present data probability

Parameters Type Default Meaning
NOISE_THRESHOLD Float 6.952 Noise level threshold in dB (See above)
PROBABILITY_THRESHOLD Float 0.600 Probability threshold for out2 (See above)

Documentation for CTKv1.1.4 - Last modified: Thu Jun 28 15:54:04 BST 2001