A noise elimination method.

The invention relates to a method for eliminating noise in an information channel by filtering a noisy signal coming through the channel.

Disturbing noise often appears in an information channel, specifically a communications channel, such as a mobile radio telephone channel. The noise may be due both to the connection and to noise around the actual source, the speaker. For example, the background noise in an automobile is heard as noise in the background of a telephone call spoken over a mobile radio telephone.

To eliminate disturbances and noise it is possible to use a plurality of methods. For example, by means of a high-pass filter it is possible to eliminate noise at low frequen­cies. However, in a signal coming to a mobile telephone from stations, the noise is within the same frequency range as is the speech signal, and so a situation is arrived at in which speech is also eliminated if attempts are made to eliminate the noise.

Endeavors have been made to eliminate the effects of back­ground noise, when speaking from automobiles, by using, for example, a second microphone. Only noise is received in this microphone placed elsewhere in the automobile, and this noise can be eliminated fromthe entire signal before the speech signal is transmitted further from the tele­phone. Only slight improvement of the signal to noise ratio has been achieved by this method, since suitable placing of the microphone is difficult.

The object of the present invention is to eliminate the problems appearing in the methods described above and to provide a method by which it is possible to eliminate noise in an information channel without suppressing the speech signal.

According to the invention the problem presented is solved by first dividing the frequency band of the channel into sub-bands, whereby the noisy information is divided into sub-signals, and by thereafter separately determining the useful value of the signal of each sub-band by comparing the signal level to a reference level, in which case a signal substantially of the reference level is defined as noise and is suppressed, whereas a varying signal, defined as a useful signal, is summated, substantially unsup­pressed, with the useful signals of the other sub-bands in order to form filtered outgoing information.

Frequency bands containing nothing but noise or a great deal of background noise can in this case be eliminated. This is possible, since a speech signal does not use the entire frequency range all the time.

The invention is described below with the help of an embo­diment example, with reference to the accompanying drawing, in which <ul id="ul0001" list-style="none"><li>Figure 1 is a block diagram of the principle of an arrange­ment with which the method according to the invention is implemented;</li><li>Figure 2 is the construction of a four-channel filter bank;</li><li>Figure 3 depicts a filter pair consisting of a low-pass filter and a high-pass filter;</li><li>Figure 4 is a block diagram of an arrangement with which the suppression of a single channel is controlled;</li><li>Figure 5 illustrates the division of a frequency band into sub-bands;</li><li>Figure 6 is a characteristic curve determining the amount of signal passage.</li></ul>

The principle of the implementation of the method is de­picted in Figure 1 as a block diagram. A signal 1 is sup­plied to the filter blocks CH1-CH4, in which the signal is divided into frequency bands B1-B4 (Figure 5). The outputs of the filters are coupled to suppressors P1-P4, by means of which it is possible to eliminate frequency bands which contain nothing but noise. Finally the bands are again summated, and the improved signal 2 is carried further.

The division into frequency bands B1, B2, B3 and Bn is implemented using the all-pass filter construction accord­ing to Figure 2. The limit frequencies of the frequency bands are f₁, f₂, f₃.

The transfer function of the all-pass filter of the first stage is A₁ = (a + z⁻¹) / (1 + az⁻¹) and respectively the transfer function of the all-pass filter of the second stage is A₂ = (b + cz⁻¹+ z⁻²) / (1 + cz⁻¹ + bz⁻²) where coefficients a, b and c are obtained from the formulae a = sin(ω/2 - π/4) / sin(ω/2 + π/4) b = (a² + α) / (1 + αa²) c = (2a(1+α)) / (1 + αa²) where ω = 2πf/f_s and α = f_s/4f_H and f_H = limit frequency of filter and f_s = sample taking frequency of filter

The filters of the first and second stages have been formed into low-pass - high-pass pairs (Figure 3) by means of which the signal is always divided into two frequency bands. Three filter stages are required in order that a division into four frequency bands will occur.

The suppression control depicted in Figure 4 is carried out in each frequency band by setting the amplification of the signal according to the following equation p = max {(S-2N)/S , C} where S = signal level N = noise level C = constant indicating the maximum suppression of the bands, in practice, for example 0.2.

If there is no signal within the frequency band being exam­ined, the signal level S and the noise level N are approxi­mately equal, and the passing amount of the frequency band is C. If the signal S is substantially greater than the noise level N, the passing amount P attains a value and the signal is not suppressed at all.

The monitoring of the signal level is carried out by recti­fying (10) and low-pass filtering (11, 12) the signal of the frequency band.

The low-pass filter 11 can be, for example, the recursive filter of the first stage Y_s(k) = A_sY_s(k-1) + (1-A_s)X_s(k) where A_s(k) = coefficient Y_s(k) = output signal X_s(k) = input signal

The noise level N within each frequency band is detected by further low-pass filtering (12) the signal coming from the level detector and by detecting (13) from this the long-­term minimum. Thus it is possible to determine whether the frequency band in question has a constant-value signal (which is noise) or a signal of continually changing value, which can be deemed to be a useful signal. The recursive filter of the first stage may again serve as the low-pass filter Y_n(k) = A_nY_n(k-1) + (1-A_n)X_n(k) where the terms stand for the same as in the previous form­ula. Referring to Figure 4, X_n(k) = Y_s(k).

The minimum is calculated from the equation N(k) = min {Y_n(k), N(k-1) + δ} where δ = constant representing the maximum increase rate of the noise level N.

Four was chosen as the number of frequency bands in the examples. An increase in their number increases the number of all-pass blocks (quadratically). From the viewpoint of practical implementation, it is hardly worth while increas­ing the number of bands. In the present application, 3 is a good choice for the number of frequency bands.

The functioning of the method has been verified by prac­tical subjective experiments.

It is evident that the method can be applied using dif­ferent circuitries. Both analog and digital circuit con­structions are usable.