Application of digital signal processing

Date: 2015-10-07; view: 524.

Noise blanking for sound. Audio signal recorded in real acoustic environments, frequently contains unwanted noise that appear owing to environmental or recording equipment. One class of noise is additive stationary noise.

Additivity means that the noise is added to the "clean" signal y [t] does not depend on him:

х[t] = у[t] + поisе[t].

Stationarity means that the properties of the noise (power spectral composition) do not change over time.

Examples of such noise can be a constant hiss of microphone or amplification equipment, electrical hum. The work of various devices that do not change the sound over time (fans, computers) can also create noise, similar to stationary. It is not stationary noise from various clicks, strikes, the rustle of wind, the noise of cars.

There is an algorithm of spectral subtraction to suppress the additive stationary noise. It consists of the following stages:

1. Decomposition of the signal using the short-term Fourier transforms (STFT) or other conversion with compact localizing of signal energy.

2. Evaluation of the noise spectrum.

3. "Subtraction" of the amplitude spectrum of the noise from - the amplitude spectrum of the signal.

4. 4. The inverse transforms STFT - the synthesis of the resulting signal.

As the filter bank it is recommended STFT with window Hannah ,

0 £ t < N, of length of 50 ms, and with the engagement factor of 75%. The amplitude of the weighting window should be scaled so that the engagement factor with the selected filter bank windows did not change the overall amplitude of the signal in the absence of processing.

Evaluation of the noise spectrum can be carried out as automatically by searching the minimum energy segments in each frequency band, or manually, by analyzing the spectrum at the time slot that the user identified as noise. Subtracting the amplitude spectrum can be obtained through the formula , that is equivalent to the following function of suppress: . Here Х[f,t] и W[f,t] – amplitude spectrums of the signal and noise, respectively, Y[f,t]=G[f,t]X[f,t] - amplitude spectrum of the resulting purified signal and k - coefficient of suppression. Phase spectrum of the purified signal is assumed to be equal to phase spectrum of the noisy signal. One of the problems of spectral subtraction method is the so-called "Musical noise". It appears because the STFT coefficients of noisy signals are statistically random, which leads to uneven of the suppression of these coefficients. As a result, the purified signal contains brief and limited in frequency bursts of energy, which the ear is perceived as "handbells" or "flowing water". In some cases, this effect is even less desirable than the original noise, which is suppressed. To suppress this artifact can be used the following methods:

- overestimation of the noise threshold (increasing k). Leads to the suppression of weak components of the desired signal, the sound is muffled.

- incomplete noise reduction (restriction G [f, t] from below by a constant, which is different from zero.) Part of the noise is remained in the signal and is partly masked "musical noise".

- Smoothing in time of the evaluations of the spectrum X [f, t]. Leads to tailing or suppression of transients (spikes in the signal: strikes, attacks musical instruments).

- Adaptive smoothing estimates of the spectrum X [f, t] (or relations X [f, t] / W [f, t]) with respect to time and frequency. The highest quality, but also time consuming.

The most common way to suppress the "musical noise" - uses a smoothing of the spectrum by time. For this to the STFT-coefficients of the original signal is applied recursive filtering by time: . Here 0 <а < 1 - constant, which controls the power smoothing.