Wideband Acoustic Echo Canceller: AEC G4

AVAILABILITY

ADT AEC G4 is available off the shelf on the following Platforms: Other configurations are available upon request.

Product Number	Platform
ADT_aec_G4_armA8	ARM Cortex-A8
ADT_aec_G4_armA9	ARM Cortex-A9
ADT_aec_G4_armA15	ARM Cortex-A15

DESCRIPTION

The Adaptive Digital Technologies acoustic echo canceller (AEC) electronically removes both direct coupling and reflected echo, enabling true full-duplex hands-free telephony in mobile phones, speakerphones, and hands-free intercoms.
The fourth-generation canceller includes a noise reduction feature, as well as the typical anti-howling, nonlinear processing, and double-talk detection found in the first generation AEC. The acoustic noise reduction feature is particularly useful during speakerphone calls where there is background noise present such as an exhaust fan, and in noisy outdoor environments. By using Adaptive Digital’s acoustic echo canceller to eliminate this unwanted echo and reverberating interference, echo-free conversation can be achieved.

FUNCTIONAL DESCRIPTION

The following figure is a simplified block diagram of the Acoustic Echo Canceller. A description follows. In the description, you will find parameter names blue, underlined, and in inside parenthesis ( parameter ). This indicates that a user-controlled parameter is involved in the process being described.

The top half of the diagram shows the receive signal path, or the signal path from the telephone network to the speaker.

The bottom half of the diagram shows the transmit signal path from the microphone toward the telephone network. The AEC cancels the echo that occurs between the speaker output and the microphone input.

The terms Rx (Receive) and Tx (Transmit) may be confusing at first because both the receive and transmit paths have inputs and outputs. The names receive and transmit are used from the point-of-view of the person at the speaker/microphone side.

The terms Rx (Receive) and Tx (Transmit) may be confusing at first because both the receive and transmit paths have inputs and outputs. The names receive and transmit are used from the point-of-view of the person at the speaker/microphone side.

The RxIn signal coming from the network is fed into the RxNLP (Receive Nonlinear Processor). The RxNLP can attenuate (MaxRxAttendB ) the received signal by a variable amount based upon the talk state (single talk vs. double-talk). This attenuation improves the overall echo attenuation.

The output of the RxNLP is fed into a noise generator, which adds noise ( maxRxNoisedBm ) to the receive signal. This noise helps the AEC converge and remain converged even when nobody is speaking. The output of the noise generator is fed both to the transmit output (TxOut) and into the bulk delay block, whose delay is controlled by bulkDelaySamples. The bulk delay block compensates for the buffering delay at the RxOut and TxIn interfaces as well as any other non-acoustic system delays in the path between RxOut and TxIn. The output of the bulk delay is fed to the the adaptive filter.

The adaptive filter estimates the echo and subtracts it from the TxIn signal to form the residual signal. The residual signal is fed to the noise reduction block. This noise reduction block removes background noise and therefore improves the signal to noise ratio of the transmit signal. Noise Reduction parameters include ( nrSTIntervalMSec , nrLTIntervalMSec , nrMaxLossdB , nrHighSNRMarkdB , nrLowSNRMarkdB .)

The output of the noise reduction block is fed into an equalizer. The equalizer is used to flatten out the frequency response of the transmit channel. This may be necessary due to the acoustics of the hands-free device and due to the characteristics of the microphone itself. The equalizer gain vs. frequency band table is provided by the user parameter ( pTxEqualizerdB10 ).

The output of the transmit equalizer is fed into the TxNLP. The TxNLP is the transmit non-linear processor. The TxNLP increases the echo attenuation by attenuating the residual by a variable amount based upon the talk state. The TxNLP is controlled by ( maxTxLossB and targetResidualLeveldBm, worstExpectedERLdB, and maxTxNLPThresholddB ). The TxNLP block also includes a comfort noise generator that is turned on or off using ( cngEnable) .

The compute gain block computes the AGC gain. The output of the TxNLP is fed into the AGC gain block, which provides gain or loss depending upon the residual signal level. The AGC is controlled by ( agcEnable ,  agcMaxGaindB , agcMaxLossdB , agcTargetLeveldBm , and agcLowSigThreshdBm .)

The output of the AGC is fed to the transmit output (TxOut) of the AEC.

USER CONTROLLED PARAMETERS (SUMMARY)

1. Sampling Rate
2. Tail Length
3. Frame Size
4. NLP Control
5. Max Noise Level
6. AGC Control
7. Equalizer Control
8. Noise Reduction Control

SPECIFICATION

ARM® DEVICES

Cortex-A8 / A9 / A15

CPU UTILIZATION & MEMORY REQUIREMENTS
All Memory usage is given in units of byte.

Sampling Rate (Hz)	Tail Length (msec)	MIPS	Program Memory	Data	Scratch	Per-Channel	Stack
8000	32	31.5	113K	6.7K	0	26K	20K
8000	64	38	113K	6.7K	0	41K	20K
16000	32	71.5	113K	6.7K	0	41K	20K
16000	64	94	113K	6.7K	0	77K	20K

Specifications are approximate. Exact per-channel numbers can be obtained by using API functions AECG4_alloc or AECG4_staticAllocHelper.