Perception and Automated Assessment of Recorded Audio Quality, Especially User Generated Content

我们中的许多人现在都随身携带着能够录音的技术，无论是用数码相机录下孩子第一次音乐会的声音，还是用手机录下一个恶作剧。如今，这种用户生成的内容有很多出口。仅去年一年，就有1300万小时的视频被上传到YouTube。即使是专业的广播公司也依赖这些镜头。主流新闻简报经常使用业余爱好者拍摄的戏剧性事件片段（例如协和飞机坠毁），而一些电视节目，如“粗鲁管道”，则完全由用户生成的内容组成。然而，用户生成内容的声音质量通常很差：失真、嘈杂、语音混乱或音乐模糊。我们的兴趣在于记录不良的原因，特别是在声源和从麦克风发出的电子信号之间发生了什么。典型的问题包括：用麦克风讲话；由于剪切造成的言语扭曲；风噪声和麦克风处理噪声。我们对自己录制的音频以及伴随视频的音轨感兴趣。我们希望提高录音质量，以便更多用户生成的音频可以被广泛使用和创造性地重复使用。为了做到这一点，我们将了解如何感知录制错误，因为噪音和失真如何影响许多声音的音频质量感知尚不清楚。我们将开发算法来自动评估劣质录音的音频质量。一种评估录制音频质量的方法有许多潜在的用途。当广播机构收到媒体时，无论是业余还是专业人士提交的，都可以进行快速的质量评估，以确定声音是否符合广播质量，而无需耗时的试镜。在互联网上搜索声音以进行创造性再利用是一项令人沮丧的活动，因为很难找到录音，而且找到的录音通常质量很差。音频质量评估方法将使标记和搜索声音文件的内容和质量成为可能。更好的是，在录制时使用音频质量评级来尝试提高捕获声音的质量是可能的。在录音设备上显示一个简单的警告会给纠正错误的机会（当有人被录音时，一个警示灯）。此外，音频质量评级可用于生产自动纠正常见录音错误的设备。这项研究的中期目标是开发这样的算法来纠正常见的录音错误，然而，先决条件是一种可以评估音频质量的方法。这就是这个计划的重点。

Many of us now carry around technologies which allow us to record sound, whether that is the sound of our child's first music concert on a digital camera or a recording of a practical joke on a mobile phone. Nowadays, there are many outlets for this user generated content. Last year alone, 13 million hours of video was uploaded to YouTube. Even professional broadcasters rely on this footage. Mainstream news bulletins regularly use amateur footage of dramatic events (e.g. Concorde crashing) and some TV programmes such as Rude Tube are entirely made up of user generated content.However, the production quality of the sound on user-generated content is often very poor: distorted, noisy, with garbled speech or indistinct music. Our interest lies in the causes of the poor recording, especially what happens between the sound source and the electronic signal emerging from the microphone. Typical problems include: speaking off microphone; distorted speech due to clipping; wind noise and microphone handling noise. We are interested in audio recorded on its own, as well as soundtracks accompanying videos.We want to improve the recording quality so that more user-generated audio can be widely used and re-used creatively. To do this we will develop an understanding of how recording errors are perceived as it is unclear how noise and distortion affects the perception of the audio quality for many sounds. We will develop algorithms for automatically evaluating audio quality from the poor recording.A method for evaluating recorded audio quality has many potential uses. When media is received by a broadcast organisation, whether submitted by an amateur or professional, a rapid quality assessment could determine whether the sound is of broadcast quality without time consuming auditioning. Searching for sounds on the Internet for creative re-use is a frustrating activity as it is difficult to find recordings, and those that are found are often of poor quality. An audio quality assessment method would make it possible to tag and search sound files for content and quality.Even better, it would be possible to use the audio quality rating at the time of recording to try and improve the quality of the captured sound. A simple warning displayed on the recording device would give an opportunity to correct mistakes (a warning light when someone is being recorded off-mic). Furthermore, a rating of audio quality could be used to produce devices which automatically correct common recording errors. The medium term aim of this research is to develop such algorithms to correct common recording errors, however, a pre-requisite is a method by which the quality of audio can be evaluated. And so that is the focus of this proposed project.