On acoustic-echo cancellation implementation with multiple cascadable adaptive FIR filter chips

Abstract

The implementation of an acoustic-echo canceler for a speakerphone using multiple cascadable adaptive FIR filter IC's (Motorola's DSP56200) is described. Ideally, canceling acoustic echoes requires a delay line which provides a delay equal to the impulse response time of the echo. It is shown that it is feasible to use the 24 bit coefficient precision of the DSP56200 to reduce the length of the delay line. Simulation results show that with only 256 taps an Echo Return Loss Enhancement (ERLE) > 10 dB can be achieved on an echo lo00 samples in length. L IntroductionIn recent years then has been a growing interest in acoustic-echo cancellation for speakerphone, acousticconference, and video-conference applications. Of the many phenomena which affect the speech quality in such applications, (the manifestations of) acoustic echoes are dominant. One of the most important uses of adaptive signal processing is canceling echoes. In speakerphone applications there are two echoes which must be canceled -1) acoustic echoes due to multi-path room reverberations and, 2) line echoes due to impedance discontinuities in telephone lines. In the case of line echoes, which are particularly noticeable when making long distance calls, the theory has been well documented [1,2]. This paper discusses some aspects of implementing acoustic echo cancelers.The following section describes the acoustic-echo canceler model which was assumed for the simulations and suggests performance criteria. In section I11 a brief overview of the key points of the echo cancellation algorithm is presented.Section IV deals with the results of computer simulation m s . The hardware set-up which is used to verify the results of the simulation is described in section V. The correspondence between theory and experiment will be reported at the ICASPP-89 conference in May. IL Acoustic-Echo Canceler ModelThe assumed acoustic-echo canceler model is shown in Fig. 1. Typically, the user of a speakerphone adjusts the speaker gain, G I so that the far-end speaker sounds as loud as the near-end speaker hears himself. As a result the round trip gain in the system ends up being on the order of 0 dB. This, of course, is conducive to oscillation (howling) if the acoustic echo is significant. Depending on the characteristics of the acoustic chamber the acoustic echo may be sufficiently strong, such that loss must be i n d u c e d into the echo path. The amount of loss inaoduced is termed the Echo R e m Loss Enhancement (ERLE) defined as [3] : where €@(k)'] and €[e(&)'] are the expected values of microphone input signal power and uncanceled echo signal power, respectively, as shown in Fig. 1. A goal for ERLE is at least 10 dE3 with > 20 dB being preferable.Due to advances in CMOS process technology, inexpensive adaptive digital filters are readily available. The DSP56200 represents such a realization. This part, which is described in more detail in section V, offers 256 taps and 24bit coefficients. As is shown in the following section, ERLE is a function of many parameters inclu...

Two Microphone Noise Reduction for Speech Using VLSI LMS Adaptive Filters

“…The coefficient word size is important to achieve effective cancellation and in certain applications the difference between 16 and 24 bit coefficients can be as much as 35 dB. ref [9]. Another iniportant feature, which is often not implemented due to computation requirements, is the facility for coefficient leakage.…”