Sub-Interval and Feed forward Techniques to Improve Signal Quality

Noise on the signal line has been commonplace, and many researchers have been addressing the techniques to remove like-spectra noise from control signal. Degraded signal to noise can affect control system performance. Application of such schemes is wide ranging, including improving recognition accuracy of limited vocabulary speech interactive systems, improving speech based password access, and reducing noise pollution in many sensor and tracking applications – like radar. The problem is significant, as the characteristics of the noise sources and of the environment are often time varying. The frequency content, amplitude, phase and velocity of the undesired noise are non-stationary, and of similar spectra as the desired signal. In this work, we extend these techniques using two novel approaches, subinterval spectral subtraction techniques for momentarystationary noise and, feed-forward techniques for nonstationary noise. We show that it is possible to improve degraded signal to noise by up to 10dB, and thereby improve control system performance (in our example for key-word recognition) from 40 – 50% to better than 90%.