The problem of acoustic echo cancellation can be best explained by referring to Figure 1 which depicts the scenario that arises in teleconferencing applications. The speech signal from the far-end speaker, received through a communication channel, is broadcast by a loudspeaker in a room. The echo from the loud speaker is subsequently picked up by the microphone. This echo must be cancelled to prevent its feedback to the far-end speaker. The microphone also picks up the near-end speaker’s speech and possible background noise, which may exist in the room. An adaptive transversal filter with sufficient length is used to model the acoustics of the room. A replica of the loudspeaker echo is then obtained and subtracted from the microphone signal prior to transmission.
The problem of acoustic echo cancellation can also be posed as one of system modelling. The main challenge here is that the echo paths are spread over a relatively long length in time. For typical office rooms, echoes in the range of 100 to 250 ms spread are common. For a sampling rate of 8 kHz, this would mean 800 to 2000 taps! Thus, the main problem of acoustic echo cancellation is that of realising very long adaptive filters. In addition, since speech is a low-pass signal, it becomes necessary to use special algorithms to ensure fast adaptation of the echo canceller.
Research work to design algorithms that are appropriate for implementation of acoustic echo cancellers, both in hardware and software, have been carried out in the past three years. In particular, the use of frequency domain adaptive filtering algorithms where the fast Fourier transform (FFT) is used to reduce the computational complexity of long adaptive filters have been extensively studied. To deal with the problem of long delay (latency) of frequency domain adaptive filter (a problem that is inherent in such filters), it had been previously proposed that the adaptive filter may be partitioned into a few smaller-sized filters that can be implemented in parallel in the frequency domain. Although this structure was proposed about a decade ago, nobody had noted that it has problems with convergence. We recently found that slow convergence is a serious problem in the partitioned frequency domain adaptive filters. We have found that this problem arises mainly because of 50% overlap in the successive partitions. Thus, reduction of such overlap was expected to solve this problem. A detailed study of a new implementation of the partitioned frequency domain adaptive filters, with reduced overlap among successive partitions, was carried out and shown to perform better than the previous implementations. This study includes theoretical analysis, computer simulations, and also real time realisation of acoustic echo cancellers. A patent revealing the results of this research has been filed recently. This patent discusses in part, a real-time implementation of a 2000 tap acoustic echo canceller that has been realised on a Pentium 166 MHz personal computer. This implementation takes up less than 50% of the CPU time of the PC.
In other research, algorithms for hardware realisation of acoustic echo cancellers have been considered. In this regard, a new algorithm with very good convergence behaviour has been proposed and published in the IEEE Transactions on Signal Processing. A funding of $105k has also been obtained from NSTB to realise the proposed algorithm in hardware (custom chip). Research work to come-up with a hardware realisation of the proposed algorithm is underway and is expected to be completed by July 1998.
Figure 1. Application of acoustic echo cancellation in a teleconferencing scenario.
