Dereverberation parameter estimation device and method, dereverberation/echo-cancellation parameter estimation device, dereverberation device, dereverberation/echo-cancellation device, and dereverberation device online conferencing system

摘要

In a conventional dereverberation system, when there is a fluctuating reverberation component, it has been difficult to determine, with high accuracy, a linear dereverberation filter for removing a non-fluctuating reverberation component. An algorithm integrating a dereverberation system using a linear filter and a dereverberation system using a non-linear filter includes the function of measuring the amount of fluctuation in transfer function in a latter-stage non-linear filter over time, and controls the strength of the non-linear filter over time based on the function. In this configuration, a strong non-linear process is implemented only when the fluctuation in transfer function is large, whereby distortion in speech components can be minimized. Further, by feeding the amount of fluctuation in transfer function over time by the function back to the estimation of a previous-stage linear filter, and thus reducing the weight for a time band in which the transfer function of a linear filter estimated value fluctuates, the influence causing a decrease in estimation accuracy of the transfer function can be reduced.

主权项

1. A dereverberation parameter estimation device comprising:
a memory storing a parameter of a linear filter for removing a non-fluctuating reverberation component contained in a speech input signal, and a parameter of a non-linear filter for removing a fluctuating reverberation component contained in the speech input signal; and a processor that estimates a dereverberation parameter for acquiring a direct sound by removing the reverberation components contained in the speech input signal, and that stores the dereverberation parameter in the memory as the parameter of the linear filter and the parameter of the non-linear filter, wherein the processor is configured to: read the parameter of the linear filter from the memory, and generates an estimated value of a linear dereverberation signal by removing the non-fluctuating reverberation component contained in the speech input signal using the linear filter; read from the memory the parameter of the non-linear filter, and generates estimated values of the fluctuating reverberation component and the direct sound component contained in the estimated value of the linear dereverberation signal using the non-linear filter; based on the estimated values of the fluctuating reverberation component and the direct sound, update the parameters of the variation reverberation component and the direct sound component constituting the parameter of the non-linear filter; and successively update the parameter of the linear filter based on the parameters of the updated fluctuating reverberation component and direct sound component.