Linear predictive echo canceller integrated with relp vocoder.

(74) Agent

Hughes Network Systems Inc., 11717 Exploration Lane, Germantown, MD 20874, US The Titan Corporation, 3033 Science Park Road, San Diego, CA 92121, US David Tzat Kin Wang, San Diego, CA, US

Philip John Wilson, San Diego, CA, US

Hennk Levkowetz, JK Thorsens Patenbureau AS, Oslo (54) Designation Device for suppressing an echo signal from a transmitter channel in two-way communication networks (56) Publications IEEE-IECEJ International Conference on Acoustics, Speech and Signal

Processing, Vol.2, 7-11. April 1986, Tokyo PJ Wilson et al .: An integrated voice codec and echo canceller implemented in a single DSP processor, pp. 1333 - 1336 (57) Summary

A linear predictive echo suppressor is performed in concert with a RELP vocoder in a two-way communication network where one. received digitized speech signal (25) synt hesized by the RELP vocoder (62) from a residual signal (72) received with linear prediction coefficients (71), echoes a transmitter channel (29) and is thereby combined with an incoming digitized voice signal (28) which is to be analyzed using a RELP vocoder (60) to be output as a transmitter signal. The echo suppressor is adapted to suppress the echo signal in the transmitter channel by generating a foreground transmitter signal (38) derived by subtracting an assumed foreground value (36) of the echo signal from the combined signal (29), and applying this foreground transmitter signal (38). on the transmitter channel (30) for analysis using the RELP vocoder (60). In order to produce the foreground transmitter signal (38), the echo suppressor in a foreground filter (10) filters the received digitized speech signal (25) with adaptation coefficients (34), thereby deriving the estimated foreground value 36 of the echo signal. Further, the foreground (17) foreground transmitter signal is applied using the received linear prediction coefficients (71), and the adaptation coefficients (34) are generated (19) by cross-correlating the received residual signal (72) with the off-white foreground transmitter signal (44), the product of said cross correlation is added to the adaptation coefficients (45) used to derive the estimated foreground value (35) of the echo signal.

<img file="NO174690B_D0001.tif" />

The present invention relates to echo suppression in a two-way communication network and relates in particular to echo suppression in a digitized speech communication system and containing a digital voice codec, such as a RELP (Residual Excited Linear Prediction) vocoder.

Echoes in a communication network occur when unpowered impedance in a four-wire circuit causes a connection between data paths for transmission and reception, respectively. In two-way communication networks, it is not uncommon for a received signal to echo in a transmitter channel. In practice, e.g. telephone communication networks two-wire telephone handsets that are connected to four-wire telephone transmission lines by means of ordinary hardware components which are sometimes not properly impedance-adapted, and in such cases an echo loss in the range of 6-12 dB is common. In a communication system with input and output for speech, such as a telephone system, and where an echo-generated signal is formed, as a result, a person may hear an echo of their own speech if the circulation delay associated with the return of the echo signal is significant, such as eg. greater than 40 ms. In conventional telephone networks, echo suppression is only required for long trunk lines where significant transmission delays are present.

However, in establishing communication networks where digitized voice signals are generated and transmitted using digital voice codecs and voice compressors, the signal transmission bypasses will become more significant even over relatively short geographical distances, and when signal transmission occurs at a relatively low bit rate, such as instance. less than 16 kbit / s, echo suppression is required.

A classic echo suppressor for communication networks, such as a telephone network, is described by M. Sondi and D. Berkley in an article titled Silencing Echoes on the Telephone Network in Proc. IEEE, 1980, pages 948-963. This echo suppressor synthesizes an estimate of the echo transmitted signal and subtracts the thus estimated value from the combined signal on the transmitter channel. This estimate is made by filtering the received signal that has been echo transmitted, with fit coefficients. These matching coefficients can be generated by an algorithm where the coefficients are repeatedly updated based on a correlation of the received signal with the estimated value deviation from the combined signal on the transmitter channel.

Although this classic echo suppressor works quite well in the case of a white noise source, the converged value's convergence with the echo transmitted signal may be too slow to provide effective suppression in the case the echo signal is a highly correlated digitized speech signal. S. Yamamoto et al. describes in the article An Adaptive Echo Canceller with Linear Predictor in Trans. IECE Japan, 1979, pages 851-857, a method for overcoming this problem, in which the deduced coefficients of the received digitized speech signal for use in determining the adaptation coefficients are produced by a technique which assumes that the received speech signal is spectral in advance is flattened or whitened with linear prediction coefficients derived from the speech signal, with the aim of decorating the speech-representative components.

This technique then provides better convergence.

Echo suppressors must also be capable of correction against false matching that occurs when both transmitting and receiving voice signals occur simultaneously. This condi tion is known as double speech. K. Ochiai et al. describes in the article Echo Canceller with Two Echo Path Models, IEEE Trans.

COM-25, 1977, pages 589-55, a system for overcoming false adaptation produced by so-called double speech, where foreground and background estimates of the echo signal are generated separately and the parameter values ​​used to generate the foreground estimate are refreshed by the parameter values ​​used to generate the background estimate when a logical control circuit determines that the background estimate provides a better approximation of the echo path's transmission characteristics.

The present invention then has the object of providing an improved echo suppression device in a communication network where a received digitized voice signal echoes in a transmitter channel and is combined with an incoming transmitter signal.

According to a first aspect of the invention there is provided a means for suppressing an echo si gnal from a transmitter channel in a bidirectional communication network, the echo signal being a received digitized voice signal being thrown back on the transmitter channel and combined with an incoming transmitter signal, the device comprising :

- equipment for whitening the received digitized voice signal,

- equipment for generating a foreground transmitter signal by subtracting an estimated foreground value of the echo signal from the combined signal,

- equipment for whitening the foreground transmitter signal,

- equipment for filtering the received digitized speech signal with adaptation coefficients, thereby creating the estimated foreground value of the echo signal,

- equipment for transmitting the foreground transmitter signal to the transmitting channel, and

- equipment for deriving the adaptation coefficients by cross-correlating the whitened received digitized voice signal with the whitened foreground transmitter signal, and for adding the product of said cross-correlation to the adaptation coefficients used to create the estimated foreground signal of e.

Against this backdrop of principled prior art from the publication IEEE-IECEJ International Conference on Acoustics, Speech and Signal Processing (ICASSP 86), Tokyo, 7-11. April 1986, Volume 2, pages 1333-1336, then has this device according to the invention as a feature that further comprises:

- equipment for generating a background transmitter signal by subtracting an estimated background value of the echo signal from the combined signal,

- equipment for filtering the received digitized speech signal with said adaptation coefficients added to the cross-correlation product, thereby creating the estimated background value of the echo signal,

- equipment for updating the adaptation coefficients used to create the estimated foreground value of the echo signal with the adaptation coefficients used to create the estimated background value of the echo signal, at any time the energy of the background transmitter signal is less than. a predetermined factor relative to the energy of the foreground transmitter signal, the energy of the background transmitter signal is less than a predetermined factor relative to the energy of the combined signal, the energy of the combined signal is less than the energy of the received digitized voice signal, and the energy of the received digitized voice signal is greater than a predetermined energy level, and

- equipment for optionally delivering the background transmitter signal to the transmitter channel for transmission instead of the foreground transmitter signal, in the case where the specified conditions exist to update the adaptation coefficients used to estimate the foreground value of the echo signal.

According to a second aspect of the invention there is provided a means for suppressing an echo signal from a transmitter channel in a bidirectional communication network wherein a received digitized speech signal synthesized by a RELP vocoder based on a residual signal received together with linear prediction coefficients, is thrown back on the transmitter channel and combined with an incoming digitized transmitter voice signal to be analyzed using a broadcast RELP vocoder, the device comprising:

- equipment for generating a foreground transmitter signal by subtracting an estimated foreground value of the echo signal from the combined signal,

- equipment for whitening the foreground transmitter signal,

- equipment for filtering the received digitized speech signal with adaptation coefficients, thereby creating the estimated foreground value of the echo signal,

- equipment for deriving the adaptation coefficients by cross-correlating the received residual signal with the whitened foreground transmitter signal, and for adding the product of said cross correlation to the adaptation coefficients used to create the estimated foreground value of the echo signal, and

- equipment for delivering the foreground transmitter signal to the transmitter channel for RELP analysis.

Against this background of principled prior art also from the publication IEEE-IECEJ International Conference on Acoustics, Speech and Signal Processing (ICASSP 86), Tokyo, 7-11. April 1986, Volume 2, pages 1333-1336, then has this device according to the invention as a feature that further comprises:

- equipment for generating a background transmitter signal by subtracting an estimated background value of the echo signal from the combined signal,

- equipment for filtering the received digitized speech signal with said adaptation coefficients added to the cross-correlation product, thereby creating the estimated background value of the echo signal,

- equipment for updating the adaptation coefficients used to create the estimated foreground value of the echo signal with the adaptation coefficients used to create the estimated background value of the echo signal, at any time the energy of the background transmitter signal is less than a predetermined factor of energy relative to the energy signal , the energy of the background transmitter signal is less than a predetermined factor relative to the energy of the combined signal, the energy of the combined signal is less than the energy of the received digitized speech signal, and the energy of the received digitized speech signal is greater than a predetermined energy level, and

- equipment for optionally delivering the background transmitter signal to the transmitter channel for transmission instead of the foreground transmitter signal, in the case where the specified conditions exist to update the adaptation coefficients used to estimate the foreground value of the echo signal.

Contrary to the first aspect of the invention, described above, utilization of the residual signal received by the RELP synthesizer makes it unnecessary to pre-whiten the received digitized speech signal synthesized by the RELP vocoder, and the received linear prediction coefficients may suitably is used to whiten the foreground transmitter signal, which is utilized during the derivation of the fit coefficients.

The invention will now be further described by way of examples of preferred embodiments and with reference to the accompanying drawings, in which:

FIG. 1 is a functional block diagram of a preferred embodiment of the echo suppressor of the present invention; and FIG. 2 is a functional block diagram of a preferred embodiment of an echo suppressor of the present invention and adapted use in a communication network comprising a RELP vocoder.

In their preferred embodiments, the echo suppression system of the present invention is created as functional units in a digital signal processor, such as a Texas Instruments Model TMS32020 type processor. From FIG. 1, showing an embodiment for general purpose, it appears that the echo suppressor comprises a foreground filter 10, a background filter 11, a first subtraction unit 12, a second subtraction unit 13, a first concreting unit 14, a second concreting unit 15, a first inverse filter 16, a second inverse filter 17, an LPC (Linear Predictive Coefficients) analysis unit 18, an adaptation unit 19, a noise level assessment unit 20, a noise generator 21 and a logic unit 22 for residual echo suppression and transmission / update selections.

The echo suppressor is inserted into a bidirectional communication network such that a received digitized voice signal enters across an input terminal 24 of a reception channel 25 and is output from an output terminal 26, and a transmitted digitized voice signal enters an input terminal 28 of a transmitter channel 29 an output terminal 30. The echo suppressor is capable of suppressing signals received on the receiver channel 28 and over an echo propagation path 32 echoing in the transmitter channel 29, thereby being combined with an incoming transmitter signal on the input terminal 28.

The foreground filter 10 generates an estimated foreground value of the received digitized voice signal echo on the transmitter channel 29 by filtering the received signal with a set of n updated adaptation coefficients UC, 34 produced by the adaptation unit 19 when certain conditions exist as determined by the logic unit 22. These conditions will be discussed below. The estimated foreground value 36 of the echo is transmitted to the subtraction unit 12, which subtracts the estimated foreground value 36 of the echo from the combined signal on the transmitter channel 29 to thereby create a foreground transmitter signal 38. This foreground transmitter signal, if present, represents an aberration signal difference representation the echo of the received signal on the transmitter channel 29 and the estimated foreground value 36 of the echo.

The background filter 11 generates an estimated background value of the received digitized voice signal echo on the transmitter channel 29 by filtering the received signal with a set of n updated adaptation coefficients UC, 34, which is continuously generated by the adapter 19. The estimated background value 40 of the echo is transmitted to the subtraction unit 13, which subtracts the estimated background value 40 of the echo from the combined signal on the transmitter channel 29 to thereby create a background transmitter signal 42. This background transmitter signal 42 actually contains an deviation signal component representing the received signal echoing on transmitter channel 29 and the estimated background value 40 of the echo.

Both the foreground filter 10 and the background filter 11 are final impulse reaction filters for the purpose of forming a stable system.

The adapter unit 19 produces each set of n adaptation coefficients UC, 34 by cross-correlating a pre-whitened received digitized speech signal R, 43 with a pre-whitened foreground transmitter signal T FORE, 44. The product of this cross correlation is added to the adaptation coefficients FC 45 used by the foreground filter 10 to produce the estimated foreground echo value 36 subtracted from the combined signal on transmitter channel 29, thereby generating the foreground transmitter signal 38 as the cross correlation product. The adaptation coefficients are derived in accordance with the following algorithm, where a block of I whitened foreground transmitter signal samples T FORE is cross-correlated with a block of I whitened received digitized voice signal samples R.

IN

UC (n) = FC (n) + k Σ [T FORE (i) »R (in)] i = l where:

- UC (n) is the nth of the updated fit coefficients, - FC (n) is the nth coefficient used by the foreground filter 10, - T FORE (i) is the i th whitened foreground transmitter signal test, - R (i) is the i-whitened received digitized voice signal sample, and - k is a normalization factor proportional to the inverse value of the square sum of R (i).

Cross correlation of signal test blocks leads to better noise detection. In the preferred embodiment, 180 samples are cross-correlated in each block.

The received digitized speech signal from the receiver channel 25 is whitened by a combination of the first concreting unit 14 and the first inverse filter 16. The first concreting unit 14 pre-emphasizes the received speech signal to accentuate the high frequency speech components and produces a pre-concreted received signal 46. The LPC analyzer 18 produces linear prediction coefficients 48 based on the concreted speech signal 46. The first inverse filter 16 produces the whitened received signal R by inverse filtering of the concreted signal 46 with the linear prediction coefficients 48.

The foreground transmitter signal 38 is whitened by a combination of the second pre-emitter unit 15 and the second inverse filter 17. The second pre-emitter unit 15 concretes the foreground transmitter signal 38 to accentuate high-frequency speech components and thus emits a pre-eminent foreground transmitter signal 50. 44 by inverse filtering of the concreted foreground transmitter signal 50 with the linear prediction coefficients 48.

The noise generator 21 produces a noise signal 52 having approximately the same signal level as the background noise energy level in the transmitter channel 28 when no signal is present. The noise signal 52 is a semi-random number series produced in accordance with a noise amplification signal 54 emitted from the unit 20 which estimates the noise level, the noise level assessment unit 20 is capable of separating background noise from the call and generates the noise amplification signal 54 for activating the noise generator 21 average signal energy level on transmitter channel 29 lies between a predetermined minimum value and a predetermined largest value. This minimum value is an empirically estimated, least applicable noise amplitude, while said greatest value is an empirically estimated energy level where the signal on transmitter channel 29 becomes so high that it will be interfering. The noise amplification signal 54 represents a noise level amplitude estimated by summing a predetermined proportion of the latter estimated amplitude with the product of the average amplitude level of transmitter channel 29 when no signal is present and the complementary value of the predetermined proportion.

The residual echo suppression and transmit / update selection logic unit 22 determines when an updated set of adaptation coefficients UC is to be output to the foreground filter 10 and which of the signals to be output to the transmitter output 30 in accordance with certain prescribed conditions. A new set of updated coefficients UC is emitted to the foreground filter 10 at any time the energy of the background transmitter signal 42 is less than a predetermined factor A relative to the energy of the foreground transmitter signal 38, the energy of the background transmitter signal 42 is less than a predetermined factor B relative to the energy of the combined signal on transmitter channel 29, the energy of the combined signal on the transmitter channel 29 is less than the energy of the received digitized voice signal on the receiver channel 25, and the energy of the received digitized voice signal is greater than a predetermined level Ε. In the preferred embodiment, the factor A = 7/8, the factor B = 1 and the predetermined energy level E are about 2 '¹⁰. These energy levels are determined over each block of I signal samples. When this set of prescribed conditions is present, the logic circuit 22 'will selectively output the background transmitter signal 42 to the transmitter channel output terminal 30 for transmission rather than the foreground transmitter signal 38.

The logic circuit 22 selectively outputs the noise signal 52 to the transmitter channel output terminal 30 for transmission instead of the foreground transmitter signal 38 when the energy of the foreground transmitter signal 38 is less than a predetermined proportion of the energy of the received digitized voice signal on the receiver channel 25.

In the preferred embodiment, this predetermined proportion is 2⁷, so that the difference in energy level is about 21 dB.

The logic circuit 22 also selectively outputs the noise signal 52 to the transmit channel output terminal 30 for transmission instead of the background transmitter signal 42 when the energy of the background transmitter signal 42 is less than a predetermined proportion of the energy of the received digitized voice signal on the receiver channel 25.

In the preferred embodiment, this predetermined proportion is 2⁷, so that the difference in energy level is about 21 dB.

The logic circuit 22 outputs the combined signal on the transmitter channel 29 to the channel output terminal 30 instead of the foreground transmitter signal 38 and resets the coefficients in the foreground filter 10 to zero when the energy in the foreground transmitter signal 38 exceeds the energy of the combined signal on the transmitter channel 29 by a predetermined value. This technique suppresses any residual echo that produces oscillations. In the preferred embodiment, the predetermined value is 100%.

Referring now to FIG. 2 illustrating a preferred embodiment of the echo suppressor of the present invention in conjunction with a RELP vocoder in a bidirectional communication network wherein a received digitized speech signal synthesized by the RELP vocoder provides echo on a transmitter channel and is combined with a transmitted digitized voice input signal to be analyzed by the RELP vocoder to be broadcast. The echo suppression is the same as described with reference to FIG. 1, except that the first concrete unit 14, the first inverse filter 16 and the LPC analysis unit 18 are omitted. The same reference numerals are used to denote functional units and signals common to the embodiments of FIG. 1 and 2.

The RELP vocoder comprises a RELP analyzer 60 and a RELP synthesis unit 62, both of which are embedded in the digital signal processor constituting the echo suppressor. The RELP synthesis unit 62 comprises an unpacking and decoding unit 64, a spectral creation unit 65, a synthesis filter 66, and a dimming unit 67. A signal 69 generated by a RELP analysis unit is received and processed by the RELP synthesis unit to produce a received digitized speech signal 25. The RELP vocoder is described in more detail in international patent application with publication no. WO 86/02726, published May 9, 1986.

An assembly of the echo suppressor of the present invention with the RELP vocoder allows advantageous use of linear prediction coefficients 71 produced by the unpacking and decoding unit 64 as well as a residual signal 72 emitted from the spectral rectifier unit 65. Since the residual signal 72 is already a whitened version of the received digitized voice signal on the receiver channel 29, there is no separate need to pre-whiten the received digitized voice signal, as was the case with the embodiment of the echo suppressor shown in FIG. 1. The received linear prediction coefficients 72 are also used for inverse filtering of the concreted foreground transmitter signal, so that it is possible to eliminate the LPC analyzer 18 of FIG. 1.

In the coordinated system of FIG. 2, the adapter 19 produces each set of n adaptation coefficients UC, 34 by cross-correlating the residual signal 72 with the whitened foreground transmitter signal T FORE, 44, the product of the cross-correlation being added to the matching coefficients FC 45 used by the foreground filter to produce it 36. The above equation determines the generation coefficients under the assumption that R (i) is the last residual signal sample, instead of the one whitened received digitized speech signal sample, as was the case for the echo suppressor of FIG. 1.

The signals which in the embodiment shown in FIG. 2 is selectively emitted for transmission from the output terminal 30 of the transmitter channel, is output to the RELP analyzer 60 to be subjected to RELP analysis.

The echo suppressor of the present invention is particularly suitable for assembly with a RELP vocoder, both of which are adapted to process blocks of signals so that the integrated signal processing can be easily synchronized.