A method and device for a frequency-domain echo suppression filter

Abstract

Method for suppressing residual frequency components of a reference signal from a least-mean-squared (LMS) filtered signal carrying these residual frequency components of the reference signal, wherein the method comprises: Dividing a magnitude of the frequency domain representation of the reference signal by a magnitude of the frequency domain representation of the LMS-filtered signal to obtain a frequency domain ratio of the frequency domain representation of the reference signal to the frequency domain representation of the LMS-filtered signal; and Multiplying the frequency domain representation of the LMS-filtered signal by the frequency domain ratio of the frequency domain representation of the reference signal to the frequency domain representation of the LMS-filtered signal, to obtain a frequency range signal that has reduced residual frequency components of the reference signal.

Description

BACKGROUND

[0001] Fig. 1 illustrates an automobile or other motor vehicle 100 equipped with a hands-free cellular telephone 102 that provides two-way voice and data communication over a cellular telephone network 104. The cellular telephone 102 and the network 104 over which the cellular telephone 102 communicates provide the vehicle's occupants with the ability to continue a two-way voice conversation via a cellular telephone 105 located away from the vehicle 100.

[0002] While cellular networks are nearly ubiquitous and the communications they provide are relatively reliable, audio, or listening to a hands-free cellular phone, continues to suffer from a phenomenon known as "far-end echo." Fig. 1, voice or other audio signals 106 output from a speaker 108 are at an amplitude large enough to allow the audio signals to be heard by a driver 110 or other vehicle occupants. These audio signals are therefore inherently strong enough to be sampled or "picked up" by a hands-free microphone 112 within the vehicle.

[0003] During a hands-free call between the cellular phone 102 in the vehicle and the remote or "long-distance" cellular phone 105, audio signals detected by the hands-free microphone 112 are transmitted from the cellular phone 102 to the remote or "long-distance" cellular phone 105. Since the audio output from the speaker 108 must be loud enough for the driver 110 and vehicle occupants to hear, the audio from the speaker 108 is also picked up by the microphone 112. Audio output from the speaker 108 that originates from the long-distance cellular phone 105 is thus picked up by the microphone 112 and sent back to the long-distance cellular phone 105 as an "echo."

[0004] Far-end audio echo, i.e., audio from a distance that is picked up at the "near end" and retransmitted back to the far end, can be reduced by filtering the output signal of a hands-free telephone microphone 112 near a link, but prior art echo cancelers are not capable of completely suppressing far-end echo. A method and apparatus for suppressing the echo remaining in a signal after it has been LMS-filtered would be an improvement.

[0005] US 6859531 B1 describes how real and imaginary components of the discrete Fourier transform (DFT) of a residual echo in an acoustic echo cancellation system are estimated separately using mapping functions derived without assuming statistical independence between separate components of a near-infrared microphone input signal. The residual echo estimator calculates the real and imaginary components based on an approximation of the microphone input signal without echo. SHORT SUMMARY

[0006] Embodiments are directed to suppressing residual frequency components of a reference signal from an error signal. A magnitude of the frequency-domain representation of the reference signal is divided by a magnitude of the frequency-domain representation of the LMS-filtered representation of the error signal to obtain a frequency-domain ratio of the frequency-domain representation of the reference signal to the frequency-domain representation of the LMS-filtered signal. The frequency-domain ratio of the frequency-domain representation of the reference signal to the frequency-domain representation of the LMS-filtered signal is multiplied by the frequency-domain ratio of the frequency-domain representation of the reference signal to the frequency-domain representation of the LMS-filtered signal to obtain a frequency-domain signal having reduced residual frequency components of the reference signal. BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 is a drawing of a radio communication system consisting of a hands-free cellular telephone; Fig. 2 is a block diagram of a time-domain least-mean-squared (LMS) filter; Fig. 3 is a block diagram of a frequency domain LMS filter; Fig. 4 is a block diagram of an apparatus and method for providing frequency domain echo cancellation; Fig. 5 is a block diagram of another apparatus and method for providing frequency domain echo cancellation; Fig. 6 is a block diagram of another apparatus and method for providing frequency domain echo cancellation and providing scaling of frequency components; and Fig. Figure 7 is a block diagram of an implementation of a device to provide frequency domain echo cancellation. DETAILED DESCRIPTION

[0007] As used here, the term "adaptive filter" refers to a filter that self-tunes its transfer function according to an optimization algorithm driven by an error signal. The optimization algorithm typically operates on a signal input to the adaptive filter.

[0008] Most adaptive filters are digital filters because the optimization algorithms are complex. Digital signal processors are often used to implement a digital filter, but robust microcontrollers and microprocessors can also implement the calculations required by a digital filter. The functional means described below are therefore preferably implemented using one or more computers or processors that execute program instructions stored in one or more non-transitory memory devices accessible by one or more computers or processors.

[0009] Fig. Figure 2 is a block diagram of a time-domain adaptive filter 200 embedded as a least-mean-squared, or "LMS," filter. U(n) is an audio "reference" signal 202 received by one end 201, typically a cellular phone, not shown. The reference signal 202 is output by a conventional loudspeaker 204, but at least a portion of the audio output from the loudspeaker 204 is also picked up by a microphone 208. The audio signals picked up by the microphone 208 are retransmitted back to the end 201 unless removed or at least suppressed from the microphone's output before being modulated onto a radio frequency carrier.

[0010] In Fig. 2, the audio signals output by microphone 208 are provided or input to a subtractor 210. Subtractor 210 removes or "subtracts" from the microphone's output signal signals of frequencies determined by a time-domain adaptive filter 214, which is a filter that selectively changes its output characteristic signal 212 in response to a control signal 216. The signal 212 output by adaptive filter 214 consists of portions of input signal 202 selected or determined by control signal 216. The output signal 218 of subtractor 210 is referred to herein as an error signal 218.

[0011] Fig. 3 is a block diagram of a frequency domain least mean squared (LMS) filter 300. Its topology is similar to the time domain LMS filter shown in Fig. 1. An audio signal 302 from a far end 301, denoted as u(n), is input to a loudspeaker 304, from which audio signals 306 are emitted. Some of the signals 306 from the loudspeaker 304 are picked up by a microphone 308.

[0012] The microphone 308 has its output signal 309, which is provided by a subtractor circuit 310. The subtractor 310 subtracts an error signal or echo estimation signal 312 from the audio output signal 309. The output signal 314 of the subtractor 310 is thus substantially similar to the audio output signal 309 from the microphone 308, but already reduced by one or more frequency components or frequency bands of the audio signal 302 received from the far end 301.

[0013] In Fig. 3, the adaptive filter element or component 316 is a frequency-domain adaptive filter. It receives a frequency-domain input signal 318 output from a fast Fourier transform element 320 having an input signal 322 that receives the audio stream or signal 302 from the far end 301. Fast Fourier transforms, circuits, and methods for fabricating them are well known in the art. A description of these is omitted for brevity.

[0014] The output signal 324 of the frequency domain adaptive filter 316 is provided to an inverse fast Fourier transform 326, the output signal 328 of which is a time domain or time-varying signal forming the echo estimate signal 312, which is input to the subtractor 310.

[0015] Fig. 3 also shows a second Fast Fourier Transform 330 receiving the output signal 314 of a subtractor 310. The second Fast Fourier Transform 330 provides an input signal 332 to the frequency-domain adaptive filter 316 that is representative of the spectral content of the output signal 314 from the subtractor 310. The spectral content 332 controls the frequency-domain adaptive filter 316 to control or cause the filter 316 to produce audio-frequency spectral components in its output signal 324, which, when fed back to the time-domain and then subtracted from the time-domain signal 309 of the microphone 308, reduce or eliminate the same spectral components from the output signal 314 of the subtractor 310. The frequency domain LMS filter 300 of the Fig. 3 selectively reduces from the audio signal 309 of the microphone 308 different spectral regions or spectral components which originate from or are produced by audio signals 306 output from the loudspeaker 304 and which originate from the far end 301, such as a cell phone.

[0016] Unfortunately, audio signals 306 originating from speaker 304 and picked up by microphone 308 are not completely removed from the signal 314 output by subtractor 310. Some residual echo is often found in the audio output signal 314 from the subtractor. When this signal is sent to far end 301, a user at far end 301 will hear a portion of their voice as an echo, which can be annoying.

[0017] Fig. Figure 4 is a block diagram of a device 400 for suppressing echo in a frequency-domain signal. The device 400 includes an echo suppression filter 402 that receives a time-domain signal 404 from an LMS filter 406, which is similar to the LMS filter 300 shown in Fig. 3 is shown.

[0018] The echo cancellation filter 402 includes a frequency-domain divider 408 having two signal input terminals 410 and 412. The divider 408 receives a first input signal 410, a frequency-domain representation 414 of a "reference signal" output from a frequency-domain adaptive filter 416. The adaptive filter 416 includes a portion of the LMS filter 406. The reference signal 414 is a filtered frequency-domain estimate of an audio signal 418 received from a far end 401 of a communications network, not shown.

[0019] A second input signal 412 of divider 408 receives an error signal 420. Error signal 420 is a frequency-domain copy or representation of uplink signal 404, which is the output signal of a fast Fourier transform 421 that converts the time-domain representation of the output signal of subtractor 422 into the frequency domain. Divider 408, which is a frequency-domain divider, divides the magnitude of the frequency-domain representation of reference signal 414 by the magnitude of the frequency-domain representation of error signal 420. It produces a frequency-domain output signal 414, which is a frequency-domain ratio signal 424. The frequency domain ratio signal 424 represents components of the audio signal 418 from the far end 401 that remain in the signal 404 output from the subtractor 422.

[0020] The output signal 424 of the frequency-domain divider 408 is input to a frequency-domain-to-time-domain converter 426. The converter 426 is preferably embedded as an inverse fast Fourier transform (IFFT) 426, which produces a time-domain filter coefficient signal 428 as an output signal.

[0021] The time-domain filter coefficient signal 428 is input to a first input terminal 432 of a time-domain adaptive filter 430. The error signal 404, which is the output signal of the subtractor 424, is input to a second input terminal 434 of the time-domain adaptive filter 430.

[0022] The time-domain adaptive filter 430 multiplies the error signal 404 by the filter coefficient signal 428. This multiplication is performed by a conventional, prior art conversion of the error signal 404 by the filter coefficient signal 428. The output signal 436 of the time-domain adaptive filter 430 is thus in the time domain, i.e., a conventional audio signal 436 whose frequency components are found with no or substantially no frequency components in the audio signal 418 received by the far end 401. The audio echo, which would be evident in the time domain, is thus suppressed or eliminated by the operations in the frequency domain.

[0023] Those skilled in the art will recognize that it is computationally difficult to perform a conversion of the error signal 404 by the filter coefficient signal 428. It is simpler and faster to perform such an operation in the frequency domain.

[0024] Fig. Figure 5 illustrates a second and preferred embodiment of a method and apparatus for frequency-domain echo cancellation. An echo cancellation filter 502 is configured and arranged to operate in the frequency domain. The echo cancellation filter 502 has three inputs. A first input 504 receives a frequency-domain reference signal 506 from a frequency-domain adaptive filter 508, which forms part of a frequency-domain LMS filter 510.

[0025] A second input 512 of the echo cancellation filter 502 receives an error signal 514, which is a frequency-domain representation of the audio signal 516 output from a time-domain subtractor circuit 518. The error signal 514 is obtained by performing a fast Fourier transform 520 on the error signal 516. The signals received at two inputs 504 and 512 are provided to a frequency-domain divider 513, whose output signal 515 is a frequency-domain ratio of the signals received from the two inputs 504 and 512.

[0026] A third input 522 to the echo cancellation filter 502 receives the time-domain representation of the error signal. A fast Fourier transform 524 within the echo cancellation filter 52, which may also be external to the filter 502, such as between the echo cancellation filter 502 and the LMS filter 510, converts the time-domain error signal 542 to the frequency domain and provides the frequency-domain representation of the error signal 526.

[0027] The frequency domain multiplier 526 has two inputs: 528 and 530. A first input 528 receives the frequency domain representation of the error signal 516. The second input 530 receives the output signal of a frequency domain divider 513. The divider 513 calculates a frequency domain division or quotient of the magnitude of the frequency domain representation of the reference signal by the magnitude of the frequency domain representation of the error signal 514. The output of the divider 513 is a frequency domain signal 515, which is a representation of the ratio of those two frequency domain signals.The output signal 515 is multiplied in the multiplier 526 by the frequency domain representation of the error signal 522 to produce a frequency domain output signal 537 whose frequency components are devoid of, or substantially devoid of, the frequency components found in the original audio signal 540 received by the far end 501.

[0028] Fig. 6 is another embodiment of a method and apparatus for frequency domain echo cancellation. The apparatus shown in Fig. 6, has a frequency domain echo cancellation filter 602, which consists of a frequency domain divider 604, with two inputs 606 and 608. The device 600, which is shown in Fig. 5 is different from the device shown in Fig. 4 is shown, as well as Fig. 5 by the use of one or more optional scaling elements 610 and 612, which may be either multipliers or dividers, which may operate on frequency domain representations of the error signal 614 and / or the reference signal 612 before being input to the divider 604.

[0029] As their names suggest, the scaling elements 610 and 612 are configured to be able to multiply or divide, i.e., to "scale," one or more frequency bands or sub-ranges of the frequency spectrum found in the frequency-domain signals embedded as the error signal 614 and the reference signal 612. Their scaling factors, however, are controllable. Individual specific band frequencies can be boosted or attenuated by different amounts or factors. By scaling one or more different frequency bands differently from others, the subsequent division by the frequency-domain divider 604 can selectively boost or suppress different frequency bands, which may be required by different applications for a frequency-domain echo canceller.

[0030] Fig. 6 illustrates another scaling element 620 on the output side of the frequency-domain divider 604. This third scaling element 620, which is also optional, can selectively scale different frequency bands or elements output from the divider 604. Converting this scaled frequency-domain signal to the time domain by the inverse Fast Fourier Transform 622 generates a coefficient signal 624, which is applied to an adaptive time-domain filter 26. By changing the filter coefficient signal using one or more of the scaling elements 610, 612, and 620, the output signal of the time-domain adaptive filter 622 can be tailored to selectively suppress or enhance different components of an audio frequency spectrum.

[0031] Professionals should understand that the use of the scaling terms as described in Fig. 6 can also be used in the full frequency range method and device described in Fig. 4 and Fig. 5. The repetition and representation of the scaling elements 610, 612 and 620 is therefore not necessary in the embodiment shown in Fig. 5 is omitted for the sake of brevity.

[0032] Fig. Figure 7 illustrates an implementation of a device for performing frequency domain echo cancellation using a conventional microcontroller or microprocessor. Fig. 7, a conventional processor 702 with I / O peripherals on the same semiconductor substrate has an audio input port 704 coupled to receive audio signals 706 from a microphone 708. The microphone 708, which is a hands-free microphone, is positioned as it would be in a vehicle, such as the one used in Fig. 1, and where it would be able to receive audio signals 712 output from a loudspeaker or other audio transmitter 714. The audio signals 716 sent to the loudspeaker 714 are depicted as coming from the processor 702, but they can just as easily be received from the radio frequency receiver 718. Similarly, the audio signals 706 detected by the microphone 708 can be provided directly to the transmitter as well as the processor 702.

[0033] Processor 702 is coupled to a memory device that is non-transitory, meaning it stores and is capable of being overwritten with program instructions that can be executed by processor 702. Memory 722 is coupled to the processor via a conventional address / data / control bus 724.

[0034] Executable program instructions in memory 722 provide processor 702 with the ability to perform the functions described above. More specifically, those instructions provide the processor with the ability to generate frequency-domain representations of audio signals, divide and multiply frequency-domain representations of two or more audio signals, and convert frequency-domain representations of signals to the time domain. These instructions also provide processor 702 with the ability to suppress frequency components of a reference signal found in the error signal.

[0035] By providing a frequency domain echo cancellation filter, such as the filters used in Fig. 4, Fig. 5 and Fig.6 on the output side of an LMS frequency-domain echo canceller, improved audio quality beyond that of the prior art can be achieved. The foregoing descriptions are for illustrative purposes only. The true scope of the invention is defined by the following claims.

Claims

[1] A method for suppressing residual frequency components of a reference signal from a least-mean-squared (LMS) filtered signal carrying these residual frequency components of the reference signal, the method comprising: Dividing a magnitude of the frequency domain representation of the reference signal by a magnitude of the frequency domain representation of the LMS-filtered signal to obtain a frequency domain ratio of the frequency domain representation of the reference signal to the frequency domain representation of the LMS-filtered signal; and Multiplying the frequency domain representation of the LMS-filtered signal by the frequency domain ratio of the frequency domain representation of the reference signal to the frequency domain representation of the LMS-filtered signal, to obtain a frequency domain signal which has reduced residual frequency components of the reference signal. [2] The method of claim 1, further comprising the step of converting the frequency domain signal having reduced residual frequency components of the time domain reference signal. [3] A method according to any preceding claim, further comprising the step of scaling at least one of the following: the frequency domain representation of the reference signal, the frequency domain of the LMS-filtered signal, and the frequency domain ratio of the frequency domain representation of the reference signal to the frequency domain representation of the LMS-filtered signal. [4] A method according to any one of the preceding claims, wherein the reference signal is an audio signal carried over a radio communication network and output from an audio converter, and wherein the LMS-filtered signal is an audio signal output from an audio signal detector, the audio signal from the audio signal detector comprising the reference signal and a speech signal. [5] A method according to any preceding claim, wherein the audio transducer is a hands-free telephone speaker, and wherein the audio signal detector is the hands-free telephone microphone. [6] An apparatus for suppressing residual echo signals from a time-domain error signal output by a least-mean-squared frequency-domain filter, the apparatus comprising: a frequency domain divider configured to divide a magnitude of the frequency domain representation of a reference signal by a magnitude of the frequency domain representation of the error signal, wherein the divider is additionally configured to provide a frequency domain representation of the ratio of the frequency domain representation of the reference signal to the frequency domain representation of the error signal; and a frequency domain multiplier configured to multiply the frequency domain version error signal by the ratio of the frequency domain representation of the reference signal to the frequency domain representation of the error signal. [7] The apparatus of claim 6, further comprising a frequency domain to time domain converter configured to receive the output signal of the frequency domain multiplier and output a time domain signal substantially comprising the error signal reduced by the reference signal. [8] The device of claim 6 or 7, further comprising a scaling element configured to scale at least one of: the frequency domain representation of the reference signal, the frequency domain of the LMS-filtered signal, and the frequency domain ratio of the frequency domain representation of the reference signal to the frequency domain representation of the LMS-filtered signal. [9] Device according to one of claims 6 to 8, further comprising: an audio converter configured to receive the reference signal and generate an audible signal therefrom; and an audio signal detector configured to detect the audible signal from the audio transducer and to generate an error output signal comprising the reference signal and a speech signal. [10] Apparatus according to any one of claims 6 to 9, wherein the frequency domain divider and the frequency domain multiplier comprise: a processor; and at least one storage device which stores program instructions which, when executed, cause the processor to: to divide a magnitude of the frequency domain representation of a reference signal by a magnitude of the frequency domain representation of the error signal, wherein the divider is additionally configured to provide a frequency domain representation of the ratio of the frequency domain representation of the reference signal to the frequency domain representation of the error signal; and multiply the frequency domain version error signal by the ratio of the frequency domain representation of the reference signal to the frequency domain representation of the error signal. [11] A method of suppressing residual frequency components of a reference signal from a least-mean-squared (LMS) filtered signal carrying residual frequency components of the reference signal, the method comprising: Dividing a magnitude of the frequency domain representation of the reference signal by a magnitude of the frequency domain representation of the LMS-filtered signal to obtain a frequency domain ratio of the frequency domain representation of the reference signal to the frequency domain representation of the LMS-filtered signal; and Converting to the time domain, the frequency domain ratio of the frequency domain representation of the reference signal to the frequency domain representation of the LMS-filtered signal to produce a time domain filter coefficient signal; and Providing the time-domain filter coefficient signal to a time-domain frequency component rejection filter configured to receive the LMS-filtered signal as an input signal, and to produce an echo-suppressed output signal responsive to the time-domain filter coefficient signal. [12] The method of claim 11, wherein the step of providing the time-domain filter coefficient signal to a time-domain frequency component suppression filter further comprises converting the LMS-filtered signal by the time-domain filter coefficient signal, and wherein the echo-suppressed output signal is a time-domain signal comprising the LMS-filtered signal reduced by the reference signal. [13] The method of claim 11 or 12, further comprising the step of scaling at least one of the following: the frequency domain representation of the reference signal, the frequency domain of the LMS-filtered signal, and the frequency domain ratio of the frequency domain representation of the reference signal to the frequency domain representation of the LMS-filtered signal. [14] A method according to any one of claims 11 to 13, wherein the reference signal is an audio signal carried over a radio communication network and output from an audio converter, and wherein the LMS-filtered signal is an audio signal output from an audio signal detector, the audio signal from the audio signal detector comprising the reference signal and a speech signal. [15] The method of any one of claims 11 to 14, wherein the audio transducer is a cellular telephone loudspeaker, and wherein the audio signal detector is the cellular telephone microphone. [16] An apparatus for suppressing residual echo signals from a time-domain error signal provided by a least-mean-squared frequency-domain filter, the apparatus comprising: a frequency domain divider configured to divide a frequency domain representation of a reference signal by a frequency domain representation of the error signal, wherein the frequency domain divider is additionally configured to provide a frequency domain ratio signal; a converter configured to convert a frequency-domain signal into a time-domain signal, the converter being coupled to the frequency-domain divider and additionally configured to generate a time-domain filter coefficient filter signal as an output signal; and a time domain filter having: a first input configured to receive the error signal, a second input configured receiving the filter coefficient signal from the converter and having an output, wherein the time domain filter is configured to suppress frequency components of the reference signal that are in the error signal. [17] The apparatus of claim 16, further comprising a scaling element configured to scale at least one of: the frequency domain representation of the reference signal, the frequency domain of the LMS-filtered signal, and the frequency domain ratio of the frequency domain representation of the reference signal to the frequency domain representation of the LMS-filtered signal. [18] Apparatus according to claim 16 or 17, further comprising: an audio converter configured to receive the reference signal and to generate an audible signal therefrom; and an audio signal detector configured to detect the audible signal from the audio transducer and to generate an error signal comprising the reference signal and a speech signal. [19] Apparatus according to any one of claims 16 to 18, wherein the frequency domain divider, the converter and the frequency domain multiplier comprise: a processor; and at least one storage device which stores program instructions which, when executed, instruct the processor: to divide a magnitude of the frequency domain representation of a reference signal by a magnitude of the frequency domain representation of the error signal, wherein the divider is additionally configured to provide a frequency domain representation of the ratio of the frequency domain representation of the reference signal to the frequency domain representation of the error signal; converting a frequency-domain signal into a time-domain signal, wherein the converter is coupled to the frequency-domain divider and is additionally configured to produce, as an output signal, a time-domain filter coefficient filter signal; and multiply the frequency domain version error signal by the ratio of the frequency domain representation of the reference signal to the frequency domain representation of the error signal.

Images