Variational Bayes state space model for acoustic echo reduction and dereverberation

In this paper, we propose a simultaneous optimization technique for speech dereverberation, acoustic echo reduction, and noise reduction, which can be utilized even when an analog-to-digital (A/D) converter and a digital-to-analog (D/A) converter are not synchronized. The proposed method utilizes a state-space model in which acoustic echo reduction filters are regarded as a time-varying state-vector due to asynchrony of the A/D converter and the D/A converter. In addition to the state-space model for acoustic echo reduction filters, the proposed method utilizes an additional state-space model in which noiseless multichannel speech signals are regarded as a state vector. By using the second state-space model, we can update the dereverberation filter under noisy environments. To optimize two types of state space models, the proposed method utilizes the variational Bayes framework. Two Kalman smoother based parameter optimization stages are performed alternatively. The proposed method is evaluated by using recorded data in a real teleconferencing room. The experimental results show that the proposed method can reduce acoustic echo signal, speech reverberation, and background noise more effectively than the conventional method by authors even when the A/D converter and the D/A converter are asynchronous.