Footnotes

One study did show an improvement in human intelligibility only by severely reducing the quality of the speech [].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... dB.²

The network was trained using standard backpropagation for 35 epochs with a step size of

, followed by 5 epochs with a step size of

. The presentation order of the training patterns was randomized.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... outputs ³

See http://www.ca.defgrp.com.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... SNR ⁴

Note that $\mathrm{SNR} = \frac{\sigma^2_{y} - \sigma^2_{n}}{\sigma^2_{n}}.$

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... estimate ⁵

Spectral subtraction can also be interpreted as a spectral based noise cancellation system, where the noise reference (spectrum) is derived from an earlier segment of the noisy signal.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... approach.⁶

In general, parametric spectral estimation methods have lower variance than nonparametric methods.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... experiment ⁷

We would like to thank Dr. Ki Yong Lee for supplying us with experimental data and details on his work.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... noise ⁸

Conversely, adaptive line enhancement techniques work by removing harmonic signals that can be predicted at long time horizons. In this case, speech plays the role of the unpredictable ``noise'' [3].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... filter ⁹

This optimization approach relates to work done by Nelson [] for system identification, and to Matthews' neural approach [] to the recursive prediction error algorithm [].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... periphery ¹⁰

The standard EKF equations are also modified to reflect this windowing in the weight estimation.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... signal,¹¹

The estimation of the noise variances was performed for each window as described in [17]. This technique effectively provides an approximate ``short-term'' SNR as opposed to the histogram techniques (see Section 14.3.1), which average the noise statistics over long segments of speech.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... data ¹²

This is developed by Lee et al. [] for the linear case.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... potential.¹³

The authors wish to acknowledge Rudolph van der Merwe for his assistance with this section.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.