房间脉冲响应模拟法及混响时间盲估计应用

在构建混响语声数据集时，由于缺乏真实长混响房间脉冲响应且模拟的房间脉冲响应与真实不符，因而导致数据驱动的混响时间盲估计模型性能下降。提出了一种基于条件生成对抗网络的房间脉冲响应模拟法，该方法利用真实的房间脉冲响应训练条件生成对抗网络，可以根据指定的混响时间模拟更加真实的房间脉冲响应。使用不同方法模拟的房间脉冲响应构建训练集用于训练盲估计模型，通过声学实验评估模型性能。实验结果表明，由该方法模拟的房间脉冲响应训练的估计模型在不同信噪比下均具有最小的均方根误差且在长混响情况下显著优于其他模型。

A room impulse response generator and its application on blind reverberation time estimation

The performance of data-driven blind reverberation time estimation network degrades because real room impulse response (RIR) lacks data with long reverberation time and there is a gap between real RIRs and simulated RIRs. In this paper, a room impulse response generator based on conditional Generative Adversarial Network is introduced. Trained with real RIRs, this network can generate room impulse responses with desired reverberation time. RIRs simulated by different methods are used to train the blind reverberation time estimators. Acoustic experiments are conducted to evaluate the performance of these estimators. The experimental results show that the estimator trained with simulated RIRs generated by the proposed method has a lower root mean square error than the baseline methods in different noisy scenarios and large reverberation scenarios.