DE19729521A1 - Noise and echo suppression method especially for hands-free apparatus - Google Patents

Abstract

The method involves providing echo cancellation for the signal to be transmitted. A power density spectrum of the remaining echo signal included in the echo cancelled signal and a power density spectrum of the noise signal are estimated. The echo cancelled signal is filtered on the basis of the estimated power density spectrums for attenuating the remaining echo and the noise. Preferably, a transmission function is estimated which is equivalent to the echo cancellation.

Description

The invention relates to a method for interference and Echo cancellation in a hands-free system device having telephone device. The Erfin dung also relates to a device for off implementation of this procedure.

When operating hands-free devices results generally the problem that the language of a Participant due to the distance to the free speaker dampened and from ambient noise is superimposed. Often the reverse Practice noise a high level, for example in Trucks or in factories, so that the Si gnal / signal to noise ratio is low. On the other hand, the Language of the other participant also from that Microphone of the speakerphone added and thus delayed by the running time of the system transfer the participant back.

To reduce noise and thus to ver improvement of subjective quality and ver consistency of the speech signal to be transmitted noise reduction systems are used. This Systems preferably operate in the frequency domain  and perform based on the estimated performance density spectrum of the noise a filtering by. The estimate itself is either made rend speech breaks or continuously.

To avoid or reduce acoustic echoes become so-called echo compensation Ren or so-called level scales used. Around a sufficiently high damping and thus a very To achieve a small residual echo is an echo compass relatively high order sator required. However, this leads to a deteriorated converter limit, especially if Um in the microphone signal there is interference. The stake A level scale has the disadvantage that Signal attenuation in a transmission direction Duplex capability is deteriorating and moreover a fluctuating level of existing surroundings leads to interference in the signal to be transmitted.

The object of the invention is therefore to to provide a method and an apparatus which effective echo and noise reduction enable.

This problem is solved by a method that has the features of claim 1. Beyond that from the invention with a device solved the features of claim 16.

The inventive method provides that in Acoustic echo contained in the microphone signal attenuate any echo canceller. The  Output signal of the echo canceller is then one downstream filter supplied on the Basis of an estimated power density spec of the residual echo signal and an estimated Power density spectrum of the interference signal a fil combination for damping the residual echo and of the interference signal.

The advantage of this method or device carrying out this method in that a single filter combined a Attenuation of residual echo and interference reduction leads.

Advantageous embodiments of the invention ben from the subclaims.

The invention will now be described with reference to an embodiment game explained with reference to the single figure. This shows a block diagram of a system for combined echo attenuation and noise reduction on.

In the single figure, a device 1 is shown, which comprises a microphone 3 and a speaker 5 . Such a speakerphone 1 is part of a car phone, for example.

The hands-free device 1 also has an echo and interference signal suppression system 7 that consists of an echo canceller 9 and a filter 11 .

The echo canceller 9 , which works in the present embodiment in the time domain, is supplied with the microphone output signal y (k). The microphone output signal y (k) is composed of a speech signal s (k), an interference signal n (k) caused by ambient noise and an echo signal caused by the loudspeaker 5 d (k). The loudspeaker 5 in turn a signal sent from a distant subscriber x (k) is supplied.

The echo canceller 9 uses the distant speech signal x (k) which is fed to it to determine an estimated echo signal (k) that it subtracts from the microphone signal y (k). The echo canceller 9 thus delivers a signal e (k) which can be represented as follows:

e (k) = s (k) + n (k) + b (k),

where b (k) represents the residual echo with

b (k) = d (k) - (k).

The signal e (k) is then fed to the filter 11 , filtered and transmitted to the remote subscriber as a signal (k).

Since the function and mode of operation of the echo compensator 9 is known, a more detailed description is dispensed with.

The filter 11 now has the task of providing a signal (k) that contains no or only small portions of the residual echo signal b (k) and the Störsi signal n (k).

The filter 11 operates in the frequency domain, which is indicated in the figure by two Fourier transformation components 13 , 15 , the component 13 carrying out a fast Fourier transformation (FFT = Fast Fourier Transformation) from time to frequency range and the component 15 performing one fast Fourier transformation from the frequency domain to the time domain. The arrow 17 indicates that the filter 11 does not perform a phase shift, but only changes the amplitudes of certain frequency ranges.

The spectrum S (Ω _i ) that is supplied to module 15 is as follows:

S (Ω _i ) = H (Ω _i ) E (Ω _i ),

where E (Ω _i ) gives the discrete Fourier transform of a frame of N samples of the signal e (k) and i = 0.1. . .M-1 a frequency index with Ω _i = (i / M) .2π.

For the sole noise reduction there are some weighting rules H (Ω _i ) which only change the amplitude of the input signal and leave the phase unchanged. These weighting rules include, for example, the Wiener filter and newer methods such as the "Minimum Mean Square Error Short Time Estimator" (MMSE).

For the Wiener filter, the weighting rule for a noise reduction filter H (Ω _i ) can be represented as follows:

where R _ss (Ω _i ) is the power density spectrum of the speech signal s (k) and R _nn (Ω _i ) is the power density spectrum of the interference signal n (k).

It has now been found that due to the statistical independence of the signals s (k), n (k) and b (k), this weighting rule can be modified in such a way that a combined echo attenuation and noise reduction filter is possible. The power density spectrum R _nn (Ω _i ) is replaced by the sum of the two power density spectra R _nn (Ω _i ) and R _bb (Ω _i ), the latter representing the power density spectrum of the residual echo b (k). It gives itself the weighting rule

This weighting rule provides the optimal echo attenuation and noise reduction filters in the sense of a minimal quadratic error.

The calculation of the filter H (Ω _i ), which takes place after a time frame of N samples, for example 256 samples, requires the values of the two power density _spectra R _nn (Ω _i ) and R _bb (Ω _i ). As shown below, the weighting rule H (Ω _i ) can also be calculated on the basis of signal smoking conditions. The power density spectrum R _ss (Ω _i ) required for the weighting rule is implicitly estimated.

The modification of the weighting rule for the how ner filter is also in the same way MMSE procedure possible; here too the interference signal component as the sum of the interference signal and the Re interpreted the stereo component.

The MMSE weighting rule can be specified as a function of an a priori and an a posteriori signal-to-noise ratio (SNR). The a priori SNR is defined as

where ε (.) stands for an estimate (expected value), and the a posteriori SNR is defined as

The Vienna weighting rule can be described as follows:

The above formulas relate to a noise reduction filter. However, they can also be applied in an analogous manner to a filter for damping the residual echo by replacing N (Ω _i ) with B (Ω _i ).

The a posteriori SNR can be calculated using the spectral component E (Ω _i ) and the estimated power density spectra R _nn (Ω _i ) and R _bb (Ω _i ). The a priori SNR are calculated in a recursive procedure from the a posteriori SNR using an estimated value from the previous frame. With m as the frame index for SNR ^s _n ^(m) (Ω _i ), the estimated value is obtained

where P (X) = ½ (| X | + X). SNR ^s _b (Ω _i ) is calculated analogously with a smoothing parameter α _b .

The choice of the parameters α _n and α _b depends strongly on the characteristics of the signal component to be removed. It has proven to be advantageous to set the value α _n to 0.97 and the value α _b to 0.9.

For the combined echo attenuation and noise reduction of the filter 11 according to the invention, the different SNR expressions can be combined to form SNR ^S _{b + n} (Ω _i ) and SNR ^e _{b + n} (Ω _i ). The calculation of the SNR ^e _{b + n} (Ω _i ) value can be carried out as follows:

The value SNR ^s _{b + n} (Ω _i ) can be calculated in the manner described above by a recursive method using the value SNR ^e _{b + n} (Ω _i ). Another way of calculating the value SNR ^s _{b + n} (Ω _i ), in which the two values SNR ^e _b (Ω _i ) and SNR ^s _n (Ω _i ) can be determined, is as follows:

Because SNR ^s _b (Ω _i ) and SNR ^s _n (Ω _i ) can be calculated independently of one another, the two parameters α _b and α _{n can be} optimized. This allows the filter 11 to work in different operating modes, for example as a pure noise reduction system when there is no residual echo, that is, SNR ^s _b (Ω _i ) << 1.

As mentioned, the value of the power density spectrum R _bb (Ω) is required to calculate the filter. This value is now estimated as follows:
The starting point for the considerations for estimating a value is a transfer function F (Ω) equivalent to echo compensation,

F (Ω) = B (Ω) / D (Ω)

where D (Ω) and B (Ω) the Fouriertrans are formed of the echo and residual echo.  

F (Ω) is estimated by

where R _yy (Ω), R _ee (Ω) R _dd (e) are the power density spectra of the microphone signal y (k), the echo-compensated signal e (k) and the estimated echo (k).

The estimation of the real-valued function F ^(m) (Ω) follows frame by frame with frame length M = 1, 2, 3,. . . (typically M = 128) and with m as the current frame index. By averaging F ^(m) (Ω) in K (for example K = 8) frequency bands with the limit frequencies 0 ≦ Ω ₀ <Ω ₁ <. . . <Ω _K ≦ π the mean values F _k ^(m) , k ε {0.1,. . ., K-1} determined. The cut-off frequencies Ω ₀ and Ω _K can be adapted to the bandwidth of a telephone channel, for example.

When averaging F ^(m) (Ω) in the frequency band k, i.e. over the interval I _k = [Ω _k , Ω _{k + 1} ], only the frequencies at which

applies, for example with η = 0.01. The averaging continues only if the condition mentioned is met at least over a certain amount ψ _k of frequencies in the interval I _k . Otherwise, for example, the estimate F _k ^(m-1) from the previous frame can be used or F _k ^{(m) is set} to zero. For example, ψ _k can correspond to 25% of the interval I _k .

If Ω ₀ <0, the estimate for F ₀ ^{(m) is} also used as an estimate of the transfer function of the frequencies below Ω ₀ . The same applies to frequencies above Ω _K if Ω _K <π.

The transfer function estimated in this way

^(m) (Ω) = F _k ^(m) , Ω ∈ I _k

has a staircase shape. It can continue Lich smoothed.

R _bb ^(m) (Ω) is now calculated by ^(m) (Ω) and the auto / cross power _density spectrum R ^(m) _zd (Ω), z ε {y, e, d}, for example by

Alternatively, other functions of F ^(m) (Ω) can be used that also include a constant, for example

For the suppression of the echoes not captured by the compensator, the estimated power density spectrum of previous frames can also be included, for example by

with for example λ (Ω) = min (0.5, | ^(m) (Ω) |).

In addition to the estimated power density spectrum R _bb (Ω), the power density spectrum R _nn (Ω) is required to calculate the filter H ^(m) (Ω). The method for estimating this value is known, so that its description is omitted.

With the aid of the two estimated values for the power density spectra R _nn and R _bb , the filters for the residual echo attenuation and for the interference reduction can be calculated separately

where - as already derived - applies

and

Alternatively, for increased damping, for example

to get voted.

In order to obtain an increased residual echo damping, the calculated SNR can also be weighted, for example by

with, for example, µ ₁ = 1. In order to achieve a higher voice quality, the SNR can be limited to minimum values T _b and T _n before a combination. These values can, for example, be a function of the power density of the residual echo and / or the faults, for example with

The level of the residual interference in the output signal can be set in a targeted manner by T _n .

Filtering the echo-compensated signal e (k) the speech signal s (k) is then estimated either with the calculated filter H (Ω) in Fre frequency range or with the equivalent filter h (k) in the time domain. The device according to the invention and the method according to the invention allow it So, a residual echo attenuation and a noise reduction tion by means of a filter, the Filter parameters can be set independently and can be optimized.

Claims (17)

1. A method for suppressing noise and echo in a telephone device having a hands-free device,
in which the signal to be transmitted (y (k)) is first subjected to echo cancellation, in which a power density spectrum of the residual echo signal (b (k)) contained in the echo-compensated signal and a power density spectrum of the noise signal (n (k)) are estimated and
in which, on the basis of these estimated power density spectra (R _bb (Ω), R _nn (Ω)), the echo-compensated signal is filtered in order to dampen the residual echo and the noise. 2. The method according to claim 1, characterized in that an equivalent to echo compensation transfer function F (Ω) is estimated with
F (Ω) = B (Ω) / D (Ω),
where B (Ω) and D (Ω) are the Fourier transforms of the residual echo and the echo signal, respectively. 3. The method according to claim 1, characterized in that an equivalent to echo compensation transfer function F (Ω) is estimated with
F (Ω) = (R _yy (Ω) -R _ee (Ω) -R _dd (Ω)) / (R _yy (Ω) -R _ee (Ω) + R _dd (Ω))
where R _yy , R _ee , R _{dd are} the power density spectra of the microphone signal y (k) of the hands-free device, the echo-compensated signal e (k) and the estimated echo (k). 4. The method according to claim 3, characterized in that the real value function F ^(m) (Ω) is calculated frame by frame, with a frame length m = 1, 2, 3.. .M, preferably M = 128, where m indicates the current frame index. 5. The method according to claim 4, characterized in that by averaging the transfer function F ^(m) (Ω) in K, preferably K = 8, frequency bands with the cut-off frequencies
0 ≦ Q ₀ <Ω ₁ <. . . <Q _K ≦ π the mean values F _k ^(m) can be determined. 6. The method according to claim 5, characterized in that the cut-off frequencies Ω ₀ and Ω _{K are} adapted to the bandwidth of a telephone channel. 7. The method according to claim 6, characterized in that when averaging in the frequency band k with the frequency range I _k = [Ωk, Ω _{k + 1} ] only frequencies are considered for which:
R _dd ^(m) (Ω) ≧ ηR _yy ^(m) (Ω),
with preferably η = 0.01. 8. The method according to claim 7, characterized in that the averaging takes place only when R _dd ^(m) (Ω) ≧ ηR _yy ^(m) (Ω) over a certain amount of frequencies in the interval I _{k is} satisfied. 9. The method according to claim 8, characterized in that the function ^(m) (Ω) = F _k ^(m) , Ω ∈ I _{k is} estimated. 10. The method according to claim 9, characterized in that the function ^(m) (Ω) is smoothed in time. 11. The method according to claim 10, characterized in that R _bb ^(m) (Ω) is calculated by ^(m) (Ω) and the auto / cross power _density spectrum R _zd ^(m) (Ω), with z ∈ {y, e, d}, for example with
R _bb ^(m) (Ω) = ( ^(m) (Ω) / (1- ^(m) (Ω))) ² R _dd ^(m) (Ω). 12. The method according to claim 11, characterized in that the ge estimated power density spectrum of previous frames is included in the calculation of R _bb ^(m) (Ω), preferably with
R _bb ^(m) (Ω): = R _bb ^(m) (Ω) + λ (Ω) R _bb ^(m-1 ) (Ω). 13. The method according to claim 12, characterized in that a filter H ^(m) (Ω) for the residual echo and the noise reduction as a function of the estimated power density spectra R _bb ^(m) (Ω) and R _nn ^(m) (Ω) is calculated as
H ^(m) (Ω _i ) = SNR ^S _{n + b} ^(m) (Ω _i ) / (SNR ^S _{n + b} ^(m) (Ω _i ) +1), with
SNR ^s _{n + b} ^(m) (Ω _i ) = 1 / ((SNR ^s _b ^(m) (Ω _i )) ⁻ ¹ + (SNR ^s _n ^(m) (Ω _i )) ⁻ ¹ ),
SNR ^s _n ^(m) (Ω _i ) = (1-α _n ) P (SNR ^e _n ^(m) (Ω _i ) -1) + α _n (| H ^(m-1) (Ω _i ) E ^{(m -1)} (Ω _i ) | ² ) R _nn ^(m) (Ω _i ), and
SNR ^s _b ^(m) (Ω _i ) = (1-α _b ) P (SNR ^e _b ^(m) (Ω _i ) -1) + α _b (| H ^(m-1) (Ω _i ) E ^{(m -1)} (Ω _i ) | ² ) R _bb ^(m) (Ω _i ),
where P (x) = ½ (| x | + x). 14. The method according to claim 13, characterized in that α _n = 0.97 and α _b = 0.9. 15. Device for noise and echo sub Press in a hands-free system send telephone equipment, with an echo com pensator, characterized in that the echo compensator a filter is connected after that a combined residual echo attenuation and Störge performs noise reduction. 16. The apparatus according to claim 15, characterized in that the filter is associated with an estimation device that estimates a power density spectrum (R _bb ) of a residual echo in the output signal of the echo canceller. 17. The apparatus according to claim 16, characterized in that the filter has the transmission function
H ^(m) (Ω _i ) = SNR ^s _{n + b} ^(m) (Ω _i ) / (SNR ^s _{n + b} ^(m) (Ω _i ) +1), with
SNR ^s _{n + b} ^(m) (Ω _i ) = 1 / ((SNR ^s _b ^(m) (Ω _i )) ⁻ ¹ + (SNR ^s _n ^(m) (Ω _i )) ⁻ ¹ ),
SNR ^s _n ^(m) (Ω _i ) = (1-α _n ) P (SNR ^e _n ^(m) (Ω _i ) -1) + α _n (| H ^(m-1) (Ω _i ) E ^{(m -1)} (Ω _i ) | ² ) R _nn ^(m) (Ω _i ), and
SNR ^s _b ^(m) (Ω _i ) = (1-α _b ) P (SNR ^e _b ^(m) (Ω _i ) -1) + α _b (| H ^(m-1) (Ω _i ) E ^{(m -1)} (Ω _i ) | ² ) / R _bb ^(m) (Ω _i ),
where P (x) = ½ (| x | + x) and m is a frame index.