Noise reduction using high-resolution frequency analysis

Noise-reduction methods are being studied to aid intelligible and comfortable communications. The spectral subtraction method (SS method) and other noise-reduction methods apply the discrete Fourier transform (DFT) with a window function. However, the frequency resolution and time resolution of the DFT are a trade-off, and depend on frame length. If the frequency resolution is low, noise spectra can overlap the source signal spectrum, which makes it difficult to extract the source signal from the mixed signal. Similarly, if the time resolution is low, instantaneous variation of the frequency is not observed. To solve this problem, we have applied non-harmonic analysis (NHA), which has a high frequency resolution without the influence of frame length and the window function. Therefore, we have applied NHA as preprocessing for the noise-reduction process instead of DFT. As a basic verification of the high frequency resolution enabled by NHA, the accuracy of extracting single sinusoidal waves in a noisy environment was first investigated. The signal-to-noise ratio (SNR) was used to evaluate the quality of the recovered signal. Sinusoidal waves could be extracted using NHA with a higher accuracy than the theoretical limit of DFT. The accuracy was improved by about 10 dB or more in a low-noise (input SNR: 10 dB) environment, but was not more than about 2 dB on average in a high-noise (input SNR: -10 dB) environment. Next, NHA was applied to vowel signals in a noisy environment to investigate the possibility of analyzing speech signals. The signals recovered by applying NHA were compared with those recovered by applying DFT. The improvement in the SNR for the NHA was about 2 dB higher than DFT in a white noise environment.