Comparison between DFT-, FCT-, Wavelet-, and Lattice Filter-Based Noise Reduction for ASR

We investigate noise reduction (NR) for speech signals for automatic speech recognition (ASR). We compare four transform based noise reduction algorithms according to their influence on ASR performance. These include frequency domain based algorithms using the discrete Fourier transform (DFT), the fast chirp transform (FCT), and the discrete wavelet transform (DWT), as well as a lattice filter based algorithm. Ten different types of noise with different characteristics are used, covering a broad range from almost stationary to highly non-stationary conditions. For single speakers, on average over the different noise types, a reduction of WER of up to 5.5% is possible. The DFT, DWT, and the lattice filter based algorithms lead to an improvement up to 2% relative WER on average over all speakers and noise types. Moreover, the application of the lattice filter based algorithm to recognition of noise-free data does not cause any degradation in the recognition performance. Hence the lattice filter based NR algorithm has a potential application in the real-life ASR systems.