DE2207141A1 - CIRCUIT ARRANGEMENT FOR THE SUPPRESSION OF UNWANTED VOICE SIGNALS USING A PREDICTIVE FILTER - Google Patents

Description

Circuit arrangement for suppression unwanted speech signals using a predictive filter

The invention relates to a circuit arrangement for IT; iv3rcH * üei: iing unwanted speech signals by means of a predictive :! Filters with an arrangement for deriving an electrical Weli.2rifor.a1 ,, which represents a speech signal, an arrangement for deriving g of parameters of a speech signal, which the speech signal ■ during of the intervals of the fundamental speech wave periods a controlled-v / ir-d and an arrangement based on the derived ParameLor for dls Wi.eaergewinnung the electrical waveform responds.

In telephony and other relevant fields, it is more common for the speech of one source that a listener wishes to hear is seriously disturbed in its intelligibility by the speech of a second undesired source. Numerous proposals have already been made to reduce the effects of the second language source. These relate to a relative improvement 'of the desired speech signal by the unerwünscntö Sp-r-iehsig- nal is kept relatively unintelligible or by oils of Ener-ele

309831/0818

2207U1

unwanted signal is reduced. Regardless of the type and nature of the improvements, the typical result achieved was that the desired speech signal was now more understandable than it would have been in the complete absence of speech processing. The "Freehand ¹ " telephones exemplify this problem of mutually conflicting speech sources, since their electro-acoustic transducers (loudspeakers) represent a potential source of undesired signals on the microphone of the same speech station.

The object of the present invention is now to overcome the mentioned disadvantage of disturbing a desired speech signal by an undesired, especially with "hands-free" telephones to avoid.

For a speech processing device having an arrangement for deriving an electrical waveform representing a speech signal represents, an arrangement for deriving parameters of a speech signal that are derived from the speech signal during the Intervals of the fundamental wave period is controlled, and one

309831/0618

2207H1

Arrangement based on the derived parameters for recovery of the electrical waveform, the invention resides in the arrangement for deriving the parameters of the speech signal a delay and gain parameter of undesired speech signal values (c) during an interval which derives a substantial part of the fundamental wave period of the unwanted speech signal includes which parameters are used to recover the unwanted speech signal, and that finally a combination circuit is provided which the recovered, unwanted speech signal with a Combination (a + c) of the unwanted and one wanted (a) speech signal combined so that the desired speech signal remains.

Another embodiment of the invention relates to a circuit arrangement in which the unwanted signal is received via an incoming line in a receiving point and the Combination circuit is connected to an outgoing line for the cancellation of incoming voice signals, which in Combination with the outgoing signal can be present.

309831/0018

2207U1

Further features, advantageous configurations and developments the subject matter of the invention can be found in the subclaims.

The advantages of the invention result essentially from the following consideration:

The invention is based on the general knowledge that an undesirable Speech signal can be suppressed on the basis of its speech parameters.

US Pat. No. 3,631,520 is based on certain speech parameters. There is also provided a "predictive coding ¹¹ technique for reducing the transmission bandwidth, in which an estimate of the instantaneous value of the speech signal sample is made based on a known corresponding previous value. A difference or error signal is generated from this data and sent to a remote one The receiving station is transmitted together with certain prediction parameters

309831/0818

2207U1

entire signal from the error signal, using the prediction parameters, recreated again «

It has now been recognized that the in said US patent specified process can be redesigned to substantially eliminate a given unwanted speech signal can.

The basic concept of the invention is from the unwanted signal a delay and a gain parameter derive. These parameters control a delay and gain network through which both the desired as well as the unwanted signals are transmitted »The delay time corresponds approximately to the current duration of the fundamental dwell period of the unwanted speech signal »The gain is calculated according to one of several possible formulas, so that the delayed, unwanted signal is reduced to an amplitude level the present value of the unwanted speech signal is brought. Alternatively, a less technically complex one Execution, the gain is set equal to 1. In any case, will

309831/0818

2207U1

the output signal of the network then subtractively to the unwanted voice signal or to some component Applied signal containing the unwanted speech signal. This process can be either analog or digital Technique to be carried out.

However, it is advantageous if the method is carried out with the aid of a scanning technique in which the signal is transmitted with a Repetition frequency of, for example, 6 kHz is sampled, which means that 30-60 samples are carried out per fundamental wave period will. The number of samples in a fundamental wave period varies with the frequency of the speech fundamental wave.

In one embodiment of the invention, the speech from the loudspeaker of a hands-free telephone set, which hits the microphone of the device either directly or as a reverberation reflection, largely from the output signal of the Microphones are removed. Both the reverberation signal and the direct signal are suppressed, since the unwanted Do not change speech parameters quickly while speaking.

309831/0818

In another embodiment of the invention for example, the speech of two speakers detected by a plurality of microphones and the speech of one speaker suppressed using the determined speech parameters by combining the output signals from the microphones.

It is now evident that the transformation of the above US patent disclosed method in one aspect Filter represents a filter with a comb-shaped frequency spectrum, whose minima are at the fundamental speech frequency (and its harmonics) of the unwanted speech signal »Herein differs from the "predictive" filter of the invention from a conventional echo canceller which is only reverberant Signal and this reflection from the composite signal, which consists of the signal itself and its reflection, subtracted.

The invention is described in more detail below with reference to an exemplary embodiment illustrated by drawings. It demonstrate:

309831/0818

2207U1

1 shows the schematic block diagram of a voice transmission network with a hands-free phone and the arrangement according to the invention;

Figure 2 is a schematic block diagram of the predictive filter of the present invention;

3 shows a schematic block diagram to further illustrate the invention Forecasting;

FIGS. 4-6 are graphic representations of various characteristics of the prediction circuit;

7 and 8 show two further exemplary embodiments of the invention in a voice transmission network with hands-free telephone sets
and

Figures 9 and 10 are schematic representations of the invention for multiple speaker speech suppression in a room.

309831/0818

In the first embodiment of the invention shown in FIG. 1, a hand-free telephone set with a loudspeaker 1 and a microphone 2 in a non-reverberation-free room 3 is connected to a Sprachverarbeiiungseinrichtimg 8 according to the invention. In general, the desired speech signal comes as an input signal for the microphone 2 from the speech source 4, the nearby speaker, whose signal a mainly travels along the direct path 5. The visual reflection of the source 4 is not shown in FIG or the reflection path 7. The direct speech path of the speaker at the far end is denoted by c.

The speech processing network shown in FIG. 1 consists of a so-called prediction filter 9 which is connected to the output circuit of the microphone 2. As FIGS. 2 and c show, the filter 9 consists of two parallel branches * The first branch is a prediction circuit 11, which can be a network,

309831/0818

2207U1

ίο

which consists of a delay network 12 and an amplifier 13. The second branch is a direct shunt path. Both branches are connected to one another via a subtracter 10. The prediction circuit 11 is in a still to descriptive manner controlled by a parameter extractor 14,

which is switched on in the circuit of loudspeaker 1.

According to an exemplary embodiment of the invention, the speech processing takes place digitally. A low pass 15 with a Preference is given to a cut-off frequency of 3 kHz and a 6 kHz sampler 16 are connected in series to the output circuit of microphone 2. In a similar way, a low-pass filter 17 and a sampler 18 are connected in parallel to the input circuit of the loudspeaker 1 and further connected in series with the parameter extractor 14.

A waveform representing the speaker's signal c at the distant Represents the end is shown in FIG. There is a voice signal is redundant, i.e. the signal changes its external shape and the length of the fundamental wave period from a fundamental

309831/0818

2207Ul

wave period to the next very little, the instantaneous form can or the value of the signal c can be estimated by a linear prediction based on an earlier value of the Signal c is based.

The signal c shown in Fig. 4 is from speech in successive fundamental wave periods £, X. ₃ £, -etc. built up. Unequal amplitudes are present in the speech signals in the neighboring speech fundamental wave periods of the signal c. Thus, the gain b can be calculated (not described here) such that when applied to the sampled signal of fundamental period ί , it causes that signal to match that sampled in the next fundamental period &

do

Signal is approaching. Then, when the amplified signal of the period JJ is subtractively t combined with the signal value of the period, is then obtained substantially filtered out of the signal c "Similarly, the same result obtained when the amplified signal of the period H subtractively with the composite signal, a + c is combined in the subtracter 10.

309831/0818

2207.U1

Expressed in mathematical terms, this means that W (the amplitude of the sample η which the subtracter 10 reached via the direct path in the prediction filter 9), is subtractive with W, (the amplitude of the delayed sample value) combined, where k is the number of samples in a fundamental period. The time window over which the Parameters are evaluated, on the order of the fundamental wave period, to ensure that there is enough energy is available. A time window of 30 samples with a sampling rate of 6 kHz captured in a given fundamental period a number of samples which is between half and the total number of samples.

Since speech is only quasi-stationary during intonation, the delay parameter k and the gain parameter b must be calculated periodically. This is done with the help of the digital parameter extractor 14, in accordance with the teaching of US Pat. No. 3,631,520, which has already been mentioned carried out.

309831/0818

2207U1

As stated therein, the input speech samples are stored by the scanner 18 as "frames" of signals. The memory content is then transmitted to a computing unit of Pa ^ vmeterexfraktors 14 forms a part, and in the 30 samples for correlation; - are calculated according to the following calculated values ■ "A. relation:

η -ο

W W.

η π-j

ι? W.

ί-172

η-ο

η-

Τ / 2

(D

where η is preferably in the range of 30-60 samples can lie.

The calculated width of X. are then in a network,

which is also part of de & extracts rs 18, to the location of Peaks in order to find the largest value of X.

to determine. The value of j is found in the way

that X. is the maximum of all values of X. That particular J

The value of j is the delay parameter k, which is considered to be one of the

309831/0818

22Ό7Κ1

Parameter is transmitted to the prediction filter 9. It can be seen that k is a variable delay and that is the maximum value of X. X is. The delay parameter k

J k

is for a typical voiced speech segment in Fig. ·. 6 shown.

The gain parameter b is also used by the computing circuit calculated in the parameter extractor 14 according to the following relationship:

b - ° - ° " ⁿ ' ^k U)

^N 2

Σ ^W nk

n = o ^{n K}

The gain parameter b is transmitted to the prediction filter 9 in a similar manner.

The described calculation of the delay parameter k and gain parameter b is only one of several possibilities by analyzing the speech energy content in adjacent or substantially adjacent signal sections

309831/0818

2207H1

Parameters can be calculated which, when applied to a earlier signal section are applied, the latter in shape quite similar to the shape of the present one Will hold the signal section.

The speech transmitted to the loudspeaker 1 is therefore continuously analyzed in order to determine an optimal delay parameter from it and extract a gain factor. These parameters are updated periodically, for example every 5 ms brought up to date. If there is no incoming speech signal at loudspeaker 1, there are delay and gain zero. With an incoming signal, the calculated instantaneous Signalaus output value of the prediction circuit 11 is from the subtracted undelayed, unamplified signal sample representing the signals a + c.

The reconversion into the analog form of the signal in the output circuit of the microphone 2 takes place in a digital / Analog converter 1OA.

309831/0818

22Ö7H1

The filter shown in Fig. 3 and described above has a transfer function in Z transformation notation.

H (Z) = 1 - bZ " ^k (3)

The magnitude of the frequency dependence of a typical embodiment of the filter 9 is shown in FIG. 5, in which a prediction circuit 11 is used and T indicates the sampling period.

The frequency dependence for gain parameters b S 1 is indicated by the solid curve and for gain parameters bi 1 represented by dashed lines. Since speech is dynamic during intonation, the parameters As stated above, b and k are optimally selected and periodically, for example every 5 ms, updated to be brought.

The filtering of the input languages with the help of the calculated During intonation, this parameter leads to a suppression of up to 30 dB during the presence of voiced speech segments

309831/0818

2207U1

of the unwanted signal c and to a mean suppression of about 14 dB of the unwanted signal c.

The calculated parameters b and k do not change smoothly with respect to time. The optimum delay doubles occasionally during voiced Sprachabschiiitte, also during unvoiced speech sections, the optimal delay changed liner a white ⁺ s range, while the correlation remains relatively low. However, the calculated gains are not negligible during this voiceless speaking period. The desired speech s, which is not correlated with respect to the undesired signal Cj which is to be imierdrüclrt, is attenuated when it is transmitted through e_ii filter, which has these rapidly changing filter parameters, while under such conditions no additional suppression of the undesired speech signal takes place .

In order to avoid these difficulties, a logic circuit is built in to avoid undesired changes in the filter parameters b and k to prevent. It was found that none too

309831/0818

22Q7U1

great suppression was achieved when the correlation X was less than 0.85. Therefore, the gain becomes b for X <0, set equal to zero. This is done by a parameter control circuit 19 in FIG. 1, the b for X <0.85 equals zero. The choice of X represents a compromise between one as large as possible and a sufficiently small value so that all voiced speech segments are suppressed. the resulting suppression during intonation is unchanged, while the attenuation of a second speech signal is reduced Fig. 6 shows the prevention of the delay parameter k during a typical voiced speech segment.

Because the parameters b and k differ during voiced speech segments change relatively slowly, the prediction filter 9 is effective in removing part of the reflective Signal, as well as the direct one. Specifically that part of the reflected signal whose parameters are not too dependent on the filter parameters deviates, the amplitude is reduced.

309831/0818

2207U1

In the above explanation of the invention it was assumed that in hands-free telephones loudspeakers 1 and Microphone 2 were not arranged separately from each other. In practice, however, there will be a considerable transition period for the Wandering of the signal c along the path 6 to the microphone 2 is required. It is therefore required in the speech processing network 8 to compensate for the speaker-microphone transition time. This is done by means of the parameter delay circuit 19A, which is connected in series between the output of the parameter control and the prediction filter 9. The parameter delay circuit 19A is advantageously equipped with a setting circuit 19B for the duration of the delay, which can be used to set a delay that matches the transition time characteristics of any given hands-free Telephone set corresponds.

7 and 8 show the combination of a prediction filter with an echo canceller center clipping according to FIG German patent application P 2 156 641.2. This combination represents a possible replacement for language switching that

309831/0818

2207H1

currently for echo and feedback suppression is used.

Fig. 7 shows a network designated 50 for the removal of echo from a distant speaker in a hands-free four-wire telephone set. In this figure, like numbers indicate those components that are the same as those in FIGS. 1-3 correspond. The distant echo picked up by the microphone 2 from the loudspeaker 1 is amplitude first reduced during voiced speech segments from the speech processing network 8 in the manner explained above. Reinforcement and Delay parameters b and k of the distant speech signal are measured with respect to the received loudspeaker signal and reducing the far echo component of the microphone signal by the filtering. The remaining distant signal at the exit of the voice processing network 8 is then controlled by the echo canceller removed with center clipping.

As in the above German patent application has been described, the received signal is now used to set the clip levels

309831/0818

adjust with the help of the clipping control 22 so that they just eliminate the echo. The output signal of the digital / analog converter 10A is transmitted to the filter bank 40, which contains a plurality of band filters in the speech frequency range that are contiguous to one another. In the middle of the dippr 41, the signal in each sub-band of the filter 40 was at a level mluerigekiij.-.pt, which is determined by the clipper control 22 I. eüiimm ^T u: id the energiesii in the received signal ⁷ ; .. · within each sub-band measures. The output signal of the clipper il is filtered in the filter bank 42, which has a similar structure ν / ΐ-.ί the filter bank 40. With this exhibitor. ysp '^ r. play, the copper control 22 is also advantageously used by the. Paraineterextraktorgesieuort. Since the echo vr ί the ^ "prediction. ^ Ilter £ diminished during the initiation ^ ', .- ä, the clip levels can essentially do the same amount as the iniinisp.tion diminish-3i-t therefore in ϊ'ίμ «? eil". ·. Siaaersi ^ nai shown (dashed Leitimg} /.wischt ':;, der'i pHranittereiitr actuator 14 nr / d of the clipper control, 2 _t aas a damping dec A transfer signal to be Klippernieveausceuernng 22 \ viri:? T, which is equal to the suppression achieved by άΒΏΐ language processing network 8 is «

309831/0818

■ 2207U1

The optimal operation of the clipper level control 22 can be realized if a delay is inserted in its input path around the signal transition time already mentioned to compensate between loudspeaker 1 and microphone 2. With clipping levels reduced in this way during voiced speech sections Less near-end speech mutilation is achieved by the center clipping process. Figure 8 shows a circuit for eliminating both near end echo (Echo of the distant speaker, caused by the acoustic coupling through the room acoustics) as well as the near echo (echo of the nearby speaker, caused by an inadequate hearing loss circuit) with a hands-free 2-wire telephone. The far-end echo is eliminated from network 50 in the manner described above in connection with FIG. The near echo is eliminated by a circuit 51 similar to that on the receiving side of the local 4-wire network is switched on, as the figure shows.

An alternate method of setting the Klipper level control by the parameter extractor 14 via the

309831/0818

2207H1

Parameter delay circuit 19A is shown in circuit 51. A second prediction filter 9A is used in the circuit for attenuating the signals of the clipper level control during voiced speech signals. Therefore, the clipper levels follow the signal at the entrance of the narrowband mid-cliff rs, i.e. at the exit of the prediction filter 9A.

Another embodiment allows the suppression of the speech signal of one of two speakers in a room. Fig. shows the desired speech source 23 and an undesired speech source 24, while both speech signals are the input signal for the microphones 25 and 26. The unwanted source 24 is arranged so that time delays for the direct sound transmission to the microphones 25 and 26 are the same. In the output circuit of the microphone 25 is as FIG. 2 shows, the prediction filter 9, which has already been explained earlier. The prediction circuit 11 in Fig. 3 is of a 2-microphone parameter extractor 30 is controlled.

309831/0818

2207H1

Two signals, W from microphone 25 and W from microphone 26, enter the parameter extractor 30, in which an arithmetic circuit calculates the following arithmetic values:

1 = 0 ^W "

Y = IW 'W / C (ZW ¹² ) ¹ I ² (Σ W ² ) ¹ I ² ] (5) ^J n »0 ^{J /} η ⁿⁿ " ^J

Selects a peak elective network within the extractor then the peaks of X. and Y., and a comparator

y y

finds the largest peak that occurs in both sequences X. and Y. for the same value of j. The value of j is that Delay parameter k for the undesired speech signal which is transmitted to the prediction filter 9.

An alternative method for the extraction of the parameters is shown in FIG. Two additional microphones 27 and are arranged so that the delay times of the desired

309831/0818

22Q7H1

Speech signal 24 for direct sound transmission to the microphones 27 and 28 are the same as the time delays to the microphones 25 and 26 are. The output signals of all microphones 25 to 28 are processed by a known, non-linear processor 31 processed.

The output signal of the foreman 31 contains the undesired signal and a damped and disturbed component of the desired signal, (The output si, diale can alternatively be added to merely dampen the desired signal.) The output signal of the nicii.t -lüiearen processing :; s 31 is transmitted to the speech processing network 8, which filters out the undesired speech signal 24 in the manner already described small error in the evaluation of X unä b.

309831/081 8

Claims (4)

? 6 PATENT CLAIMS Circuit arrangement for suppressing unwanted speech signals by means of a predictive filter with an arrangement for deriving an electrical waveform representing a speech signal, an arrangement for deriving parameters of a speech signal which is controlled by the speech signal during the intervals of the fundamental wave period, and an on-order, which is responsive to the derived parameters for the recovery of the electrical waveform, characterized in that the arrangement (19, 19A; Fig. 1) for deriving the parameters of the speech signal derives a delay and gain parameter of undesired speech signals (c) during an interval, which comprises a substantial part of the fundamental wave period of the undesired speech signal, which parameters are used to recover the desired speech signal, and finally a combination circuit is provided which incorporates the recovered, undesired speech signal he combines the combination (a + c) of the undesired and a desired (a) speech signal so that the desired speech signal remains. 3 0 9 8 3 1/0818 . "2207H1 received on. O 7 C. 2. Circuit arrangement according to claim 1, characterized in that the arrangement (19, 19A; Fig. 1) is connected to an incoming line for the derivation of the parameters of the undesired voice signal values, and that the combination circuit (9; Fig. 2) is connected to an outgoing line to the reflections of the to eliminate the incoming signal on the outgoing line. 3. Circuit arrangement according to claim 2, characterized in that the incoming line with the outgoing line via an acoustic path (7; Fig. 1) a hands-free telephone set is connected, whereby the combination circuit (9; Fig. 2) the incoming signals, that are transmitted to the outgoing line can be eliminated by the acoustic coupling path. 4. Circuit arrangement according to claim 2, characterized in that the arrangement (19, 19A; Fig. 1) for the incoming signals includes a device which responds to the absence of the incoming signal with the interruption responsive to the supply of the recovered signals to the combination circuit (9). 309831/08 18