Hands-Free VoIP Terminal with Gain Control Based on Neural Network

Hands-free voice terminals have to work in various situations in which the adaptive filter (used in acoustic echo cancellation module) is not capable to sufficiently suppress the acoustic echo. Such situations include echo path change, too loud far end, speaker non-linearity, volume change and so on. In these situations residual echo has to be suppressed by the additional attenuation controlled by a gain control module. In this paper a solution for gain control by neural network was proposed with two goals: a) to suppress residual echo caused by various reasons, and b) to calculate optimal gain for sending path signal. The neural network is trained to recognize situations in which the adaptive filter may efficiently suppress the acoustic echo. In this case, gain is calculated according to the in/out signal compression curve. Otherwise, when the echo suppression is not efficient enough, the gain calculated by the neural network additionally attenuates the signal to prevent acoustic echo propagation to far end. The neural network computes the output gain based on parameters of low computational complexity and because of this the proposed algorithm is suitable for implementation on real-time, embedded platforms. The proposed algorithm was tested in a small office room in which it showed better results than similar voice-over-IP solution.