Echo canceller using short time spectral analysis

The invention relates to an arrangement for communication with a subscriber.

An arrangement for communication with a subscriber is known from US Pat. No. 4,914,692, which is coupled to a telephone network via an interface. Communication with a subscriber takes place via the telephone network. The arrangement contains a voice announcement unit, a speech recognition unit, an echo canceller and a computer for controlling the communication arrangement. The communication arrangement uses the speech recognition unit to recognize voice signals from the subscriber when a call is made and, depending on these subscriber voice signals, automatically generates voice announcements (transmission signals for the subscriber) with the aid of the voice announcement unit, which are transmitted to the subscriber via the telephone network. This results in echoes of the voice announcements caused by reflection on the communication path, which are superimposed on the subscriber voice signals. The interference of the subscriber voice signals by the echoes of the voice announcements are reduced with the aid of the echo canceller, which is set in a predetermined period between the sending of a voice announcement and a response from the subscriber, so that the speech recognition unit can identify the subscriber voice signals almost without interference. Without echo cancellation, correct recognition of subscriber voice signals is not ensured for the cases in which subscriber voice signals and voice announcements occur at the same time.

The invention has for its object to improve the communication arrangement with little technical effort to ver.

The object is achieved in that the arrangement<ul id="ul0001" list-style="dash" compact="compact"><li>Spectral analysis means for the formation of short-term spectral values ​​of received signals, which are subscriber voice signals which are temporarily superimposed on echoes from the transmit signals transmitted by the subscriber,</li><li>Echo compensation means for estimating short-term spectral values ​​of the echoes and for forming difference values ​​between the short-term spectral values ​​of the received signals and the estimated short-term spectral values ​​of the echoes and</li><li>Speech recognition means for evaluating the difference values</li></ul> contains.

With the aid of such an arrangement, correct recognition of the subscriber speech signals by the speech recognition means is also ensured if the subscriber speech signals and the transmission signals for the subscriber occur simultaneously (full duplex operation). The settings of the echo compensation means are adapted to the echoes only once for each of the time periods into which the transmitted signals and the received signals are divided in order to form short-term spectral values. The computational effort for the short-term spectral analysis is small compared to the computational effort for filtering in the time domain by using an FFT (Fast Fourier Transform). The computing effort already reduced in this way compared to realizations of the echo compensation means in the time domain, in which an adaptation of the echo compensation means is carried out for each sample value of the transmit and receive signals, is further reduced in that both the echo compensation and the speech recognition are carried out in the spectral range, whereby Reverse transformations in the time domain are superfluous.

In one embodiment of the invention, the echo compensation means are provided for forming short-term spectral values ​​of the transmit signals and for adaptively weighting these short-term spectral values ​​for estimating the echoes, an LMS algorithm in particular being used for adaptively weighting the short-term spectral values ​​of the transmit signals.

In this way, an effective estimate of the short-term spectral values ​​of the echoes is obtained, which are derived from the short-term spectral values ​​of the transmission signals. The weighting of the short-term spectral values ​​of the transmission signals is adjusted with the aid of the LMS algorithm ("Least Mean Square") so that the disturbing influence of the echoes is reduced sufficiently quickly that correct speech recognition is possible.

To further develop the invention, it is provided that means for generating the transmission signals are integrated in the arrangement.

In this embodiment, the arrangement can be implemented as an independent unit that uses the transmit and receive signals to carry out automatic communication with the subscriber. In contrast, however, an external generation of the transmission signals is also possible. In such a case, the transmission signals are supplied to the arrangement from the outside.

Voice announcements are advantageously provided as transmission signals.

The described communication arrangement is thus suitable for applications in which a dialog operation between the subscriber and the communication arrangement is required. Even in cases where the subscriber sends voice signals to the communication arrangement during the voice announcements, the recognition of these voice signals is ensured by the voice recognition means. The communication arrangement can, for example, be used by travel agencies for the automatic booking of trips.

A further possibility for the embodiment of the invention lies in the establishment of a communication path between the arrangement and the subscriber by means of a telephone network.

In this case the subscriber line of the telephone network is assigned to the communication arrangement.

For a further embodiment of the invention, the arrangement for communication with the subscriber within a motor vehicle is provided. For this purpose, it is particularly integrated in a car radio.

In this case, transmission signals of the communication arrangement and subscriber voice signals are exchanged acoustically within the motor vehicle. When integrating into a car radio, it is possible in particular to control the car radio by means of speech signals even during radio operation, for example while music is being played.

An embodiment of the invention is explained below with reference to the drawings. Show it:<dl id="dl0001" compact="compact"><dt>Fig. 1</dt><dd>a communication arrangement with an echo canceller for communication with a subscriber and</dd><dt>Fig. 2</dt><dd>a partial structure of the echo canceller.</dd></dl>

1 shows a communication arrangement 1 which is coupled to a subscriber terminal 4 of a subscriber via an interface 2 and a function block 3. Function block 3 establishes a communication path between communication arrangement 1 and subscriber terminal 4 of the subscriber. In the present exemplary embodiment, function block 3 represents a telephone network. Subscriber terminal 4 is an analog or digital telephone set.

The structure of the communication arrangement 1 is described in more detail below. It contains a voice announcement unit 5, which is used to generate transmission signals S, such as voice announcements in particular, which, after being buffered in a buffer memory 6, are sent via the interface 2 to the subscriber with the subscriber terminal 4 in order to transmit information to the subscriber. Other signals such as signal tones can also serve as transmission signals S. The communication arrangement 1 further contains spectral analysis means 7, which are coupled on the input side to the interface 2. In the present case, the spectral analysis means 7 are implemented by a filter bank, as described, for example, in "PKI Technical Messages 1/1990, pages 35 to 41". The spectral analysis means 7 are used for the short-term spectral analysis of received signals E, which are composed on the one hand of subscriber voice signals generated by the subscriber and on the other hand occasionally additionally from echoes of the transmit signals S of the voice announcement unit 5. The subscriber speech signals and the echoes overlap to the reception signals E of the spectral analysis means 7. The undesired echoes which disturb the subscriber speech signals are caused by the fact that the transmission signals S are not only transmitted in the direction of the subscriber with the subscriber terminal 4, but are also reflected in the direction of the spectral analysis means 7. The reflection of the transmission signals S can be brought about by the interface 2, by the function block 3 and also by the subscriber terminal 4.

To compensate for the echoes, the communication arrangement 1 contains an echo canceller 8 which has spectral analysis means 9 on the input side. The spectral analysis means 9 carry out a short-term spectral analysis of the transmission signal samples temporarily stored in the buffer memory 6. In the present case, the spectral values ​​supplied by the spectral analysis means 9 are to be understood as components of a spectral vector X (i). Analogously to this, the spectral values ​​supplied by the spectral analysis means 7 are components of a spectral vector Y (i). i denotes a time index and numbers the time segments into which the signals supplied to the spectral analysis means 7 and 9 are divided for short-term spectral analysis.

The spectral vector X (i) is fed to a function block 10 which carries out a weighting of the components of X (i). A function block 11 is used to adapt this weighting. For this purpose, both the spectral vector X (i) and a differential spectral vector D (i) are evaluated by the function block 11. The difference spectral vector D (i) is generated by subtracting means 12 which, from the components of the spectral vector Y (i), the corresponding components of the vector X_w(i) subtract whose components are equal to the weighted components of the spectral vector X (i). The components of the difference spectral vector D (i) are thus difference values ​​which are obtained by forming the differences between the short-term spectral values ​​of the received signals E and the estimated short-term spectral values ​​of the echoes (the components of the respective vector X_w(i)) can be determined.

The difference spectral vector D (i) is supplied as the output variable of the echo canceller 8 to a speech recognition unit 13. This carries out a pattern comparison with the difference spectral vectors D (i), which are largely speech signals which are free of the echoes, and assigns them associated speech signals. This speech recognition in the spectral range is also evident from "PKI Technical Announcements 1/1990, pages 35 to 41". A function block 14 controls the voice announcement unit 5 as a function of the voice signals recognized by the voice recognition unit 13 in such a way that transmit signals S are generated which result in a meaningful continuation or also an abort of the dialogue with the subscriber.

With the help of the echo canceller 8, the influence of the echoes on the spectral vectors Y (i) is so far by subtracting the vector X_w(i) compensates for the fact that a speech recognition by the speech recognition unit 13 has a sufficiently low error rate for the recognition of the subscriber speech signals for most applications. Compared to echo compensation in the time domain, the described arrangement reduces the computation effort by 70 to 90 percent.

The spectral analysis means 7, the echo canceller 8 and the buffer memory 6 can be easily implemented using a single signal processor. The voice announcement unit 5, the voice recognition unit 13 and the function block 14 can be implemented, for example, on a PC (personal computer).

2, the operation of the function blocks 10 and 11 and the subtracting means 12 is explained in more detail. It is assumed that the spectral analysis means 7 and 9 perform short-term spectral analyzes of length N, so that the spectral vector X (i) N components X₁ (i), X₂ (i), ..., X_N(i), the spectral vector Y (i) N components Y₁ (i), Y₂ (i), ..., Y_N(i) and the difference spectral vector D (i) N components D₁ (i), D₂ (i) and D_N(i). The vector X_w(i) consists of N components w₁ (i) * X₁ (i), w₂ (i) * X₂ (i), ..., w_N(i) * X_N(i), which are each formed by multipliers 10-1, 10-2, ..., 10-N of the function block 10 in that weight factors w₁ (i), w₂ (i), ..., w_N(i) are multiplied by the components of the spectral vector X (i). The components D₁ (i) to D_N(i) of the difference spectral vector D (i) are represented by the differences <maths id="math0001" num=""><math display="inline"><mrow><msub><mrow><mtext>Y</mtext></mrow><mrow><mtext>1</mtext></mrow></msub><msub><mrow><mtext>(i) -w</mtext></mrow><mrow><mtext>1</mtext></mrow></msub><msub><mrow><mtext>(i) * X</mtext></mrow><mrow><mtext>1</mtext></mrow></msub><msub><mrow><mtext>(i), ..., Y</mtext></mrow><mrow><mtext>N</mtext></mrow></msub><msub><mrow><mtext>(i) -w</mtext></mrow><mrow><mtext>N</mtext></mrow></msub><msub><mrow><mtext>(i) * X</mtext></mrow><mrow><mtext>N</mtext></mrow></msub><mtext>(i)</mtext></mrow></math><img file="EP0695045A2_D0001.tif" /></maths> with the help of the subtracting means 12 consisting of subtractors 12-1, 12-2, ..., 12-N. educated. The setting of the weight factors w₁ (i) to w_N(i) is carried out by the functional units 11-1, 11-2, ..., 11-N, which are combined by the functional block 11. In particular, they can be used to carry out an LMS algorithm ("Least Mean Square").

An advantageous calculation rule for the weight factors is therefore:<maths id="math0002" num=""><math display="block"><mrow><msub><mrow><mtext>w</mtext></mrow><mrow><mtext>j</mtext></mrow></msub><msub><mrow><mtext>(i) = w</mtext></mrow><mrow><mtext>j</mtext></mrow></msub><msub><mrow><mtext>(i-1) + µ * D</mtext></mrow><mrow><mtext>j</mtext></mrow></msub><msub><mrow><mtext>(i-1) * X</mtext></mrow><mrow><mtext>j</mtext></mrow></msub><mtext>(i-1), j = 1,2, ..., N.</mtext></mrow></math><img file="EP0695045A2_D0002.tif" /></maths> µ here represents an adaptation constant. If this constant is set to zero, this means that no adaptation of the weighting factors w_j(i) is done. Instead of simply multiplying components X_j(i) with the weight factors w_j(i) To improve the echo cancellation, higher-order filters can also be used.

The communication arrangement 1 described by FIGS. 1 and 2 can in particular also be used for communication with a subscriber within a motor vehicle. For this purpose, the communication arrangement is integrated, for example, in a car radio. In this case, the subscriber delivers subscriber voice signals to the interface 2 acoustically in order to control the car radio. In this example, the voice announcement unit 5 corresponds to the unit of the car radio that generates music signals. The echo canceller 8 is used here to compensate for echoes of the music signals, so that the car radio can also be controlled by voice signals from the subscriber while music is being broadcast.