Multi-Speaker Separation in Reverberant Room Using Velocity-Based Microphone

音频信号分离在电子和计算机工程领域有许多应用，如**自动语音识别、语音通信、助听器等。这些应用推动了对定向麦克风解决方案的需求，该解决方案可以只拾取所需扬声器的声音，同时抑制**来自其他扬声器的声音以及背景噪声。使用**基于压力的麦克风阵列与数字信号处理(DSP)技术相结合的传统解决方案在**空间受限的应用中不能很好地工作，其中捕获的声音包含混合声音和环境噪声。Soundskrit**正在开发一种革命性的光纤定向麦克风技术，该技术可以直接测量传入声波的**粒子速度。目前，Soundskritt研发团队缺乏信号处理**专业知识，这是基于新型**光纤麦克风在多扬声器和混响环境中获取的真实数据开发声音分离软件所需的专业知识。**在这个项目中，我们打算为Soundskritt的原型开发DSP算法，以在混响室中分离多个移动**扬声器。该项目有两个主要目标：(1)使用声强定位混响房间中的有源声源，以及(2)针对所需扬声器进行波束成形，以减弱不需要的**声音和噪声。拟议的工作分为三个阶段。第一阶段是建立一个基准系统，**通过计算机模拟研究现有解决方案的性能。第二阶段是为Soundskrit原型开发DSP算法**，以在模拟环境中实现目标，并将其与第一阶段的**解决方案进行比较。在第三阶段，将使用Soundskrit原型在几个混响**房间收集真实数据，并将对算法进行实时测试和评估。本项目所取得的研究成果**将通过一份包括所有软件代码和硬件**设计的详细技术报告转移到Soundskritt。

Audio signal separation has found many applications in electrical and computer engineering fields, such as**automatic speech recognition, voice communication, hearing aids etc. These applications have driven the need**of directional microphone solutions that can pick up only the sound from the desired speaker while suppressing**the voices from other speakers as well as the background noise. Conventional solutions using array of**pressure-based microphones combined with digital signal processing (DSP) techniques do not work well in**space-constrained applications where the captured sound contains mixed voices and ambient noises. Soundskrit**is developing a revolutionary fiber-based directional microphone technology that can directly measure the**particle velocity of an incoming sound wave. Currently, the Soundskrit R&D team lacks signal processing**expertise that is required to develop sound separation software based on the real data acquired by the new**fiber-based microphone in a multi-speaker and reverberant environment.**In this project, we intend to develop DSP algorithms for the Soundskrit's prototype to separate multiple moving**speakers in a reverberant room. The project has two main objectives: (1) localizing active sources in a**reverberant room using sound intensity, and (2) beamforming toward the desired speaker to attenuate unwanted**sounds and noise. The proposed work is divided into three phases. Phase 1 is to build a baseline system to**study the performance of the existing solutions via computer simulation. Phase 2 is to develop DSP algorithms**for the Soundskrit prototype to achieve the objectives in a simulated environment and compare it with the**solutions in Phase 1. In Phase 3, real data will be collected using Soundskrit's prototype in several reverberant**rooms, and the algorithm will be tested and evaluated in real-time. The research results achieved in this project**will be transferred to Soundskrit through a detailed technical report including all software codes and hardware**design.