Noise Cancellation Frontends for Automatic Meeting Transcription

One research direction to make computers become more active machines is to use audio input with far-field omni-directional microphones, these being required to face the acoustics problem of the meeting scenario. This topic is not only interesting from an artificial listener point-of-view, but also as one of the hottest topics in searchable media. In this paper two noise suppression methods for real meeting audio are evaluated. Because of the usage of hands-free microphones in a meeting environment, the reverberation of meeting rooms and background noise sources such as projector, computer fan have huge impact on speech quality. Two denoising methods, based on statistical Wavelet filtration and short-time Fourier transform are discussed and evaluated. Whilst the quantile technique is suitable for estimating stationary noise (beamer, computer fan), non-stationary sources (closing door, chair movement) can be handled by recursive structure and adaptive temporal gain of the statistical filtration method. Besides, frequency non-linear gain can remove colored noise effectively. The proposed noise reduction methods are evaluated with speaker localizing application applied on the meeting database of the Mistral project. The results show the pros and cons of each method applied on the meeting database.