EAGER: Approximate Inference Robust Speech Processing via Sampling

原则上，语音增强和说话人识别是相关的任务。了解说话人可以实现更好的语音增强，而干扰较少的语音可以提高识别说话人的能力。这两个任务的统计模型可以结合在一起，以产生一种共同执行这两个任务的原则性方法，然而，所产生的统计模型中的精确推理的复杂性是令人望而却步的，它必须跨越非线性特征计算。因此，为了降低复杂性而牺牲一些精确推理性能的近似推理技术是令人感兴趣的。这项早期的探索性研究拨款将Gibbs抽样近似推理技术引入到联合语音增强和说话人识别中，以便与几种变分近似推理技术进行比较。研究的最终目的是将可达到的复杂性和性能空间划分为不同的区域，以指示应使用哪些技术以及作为复杂性的函数可获得的性能。为了获得这种切分，我们将在几个大型语音语料库上进行全面的经验评估。语音增强和说话人识别技术在国防、商业和医疗技术中有多种用途。通过成功地整合这两个相互依赖的任务，语音增强和说话人识别方面的性能改进将使多种技术受益。潜在的应用包括用于控制例如使用远距离麦克风的电视机的对话系统，其中加性噪声是失真的重要来源。此外，语音增强对助听器很有用。在这些设备中，依赖于说话人的先验信息获得了提高语音可理解性的能力。

Speech enhancement and speaker recognition are, in principle, related tasks. Knowledge of the speaker can allow for better speech enhancement, while speech with less interference improves the ability to recognize the speaker. Statistical models for these two tasks can be coupled to yield a principled approach to performing both jointly, however the complexity of exact inference in the resulting statistical model, which must straddle a nonlinear feature calculation, is prohibitive. For this reason, approximate inference techniques that sacrifice some performance of exact inference for lower complexity are of interest. This Early Grant for Exploratory Research brings Gibbs sampling approximate inference techniques to bear on joint speech enhancement and speaker recognition for the sake of comparison with several variational approximate inference techniques. The ultimate aim of the research is to partition the attainable complexity and performance space into different regimes dictating which techniques should be used and the performance attainable as a function of complexity. To obtain this partitioning, a thorough empirical evaluation on several large speech corpora will be carried out.Speech enhancement and speaker recognition technology finds multiple uses in defense, commercial, and medical technologies. Multiple technologies would benefit from performance improvement in speech enhancement and speaker recognition brought about by successfully integrating these two interdependent tasks. Potential applications include dialogue systems for controlling e.g. television sets using distant microphones where additive noise is a significant source of distortion. Additionally, speech enhancement is useful for hearing aids. Speaker dependent prior information gains the ability to improve the intelligibility of speech in these devises.