Joint Noise Reduction and Bandwidth Extension via Inventory Based Speech Analysis/Resynthesis

Conference: Sprachkommunikation 2010 - 9. ITG-Fachtagung
10/06/2010 - 10/08/2010 at Bochum, Deutschland

Proceedings: Sprachkommunikation 2010

Pages: 4Language: englishTyp: PDF

Personal VDE Members are entitled to a 10% discount on this title

Xiao, Xiaoqiang; Nickel, Robert M. (Department of Electrical Engineering, Bucknell University, Lewisburg, PA 17837, USA)

Recently the authors published implementation details and performance results for a novel speech enhancement paradigm. The new approach is based on a parametric speech analysis and resynthesis scheme that is fundamentally different from most conventional and state-of-the-art denoising systems. Instead of filtering a distorted signal, a set of speech-property parameters is estimated and a new "clean" speech signal is resynthesized from a pre-recorded (clean) speech waveform inventory. The new enhancement paradigm has many potential applications, beyond the reduction of additive background noise. For example, in 2009 the authors presented a paper that utilized the same principle for joint speech denoising and low bit-rate speech encoding (1.5 kbit/sec). In this paper we are demonstrating that the approach can also be used jointly for additive noise suppression and artificial bandwidth extension (BWE). With the proposed method it is possible to reduce the high-band root-meansquare log-spectral distortion (RMS-LSD) of bandwidth extended signals to levels that are at least comparable (if not slightly superior) to those reported for other state-of-the art systems under similar noise conditions. A comparison mean opinion score test (CMOS - after ITUT recommendation P.800) also revealed that a significant reduction in background noise is achieved (albeit at the expense of a slight signal distortion). An advantage of the proposed method is that through its processing paradigm speech coding, enhancement, and bandwidth extension can, at least in theory, be performed jointly.