Noise Reduction

When a speech communication device is used in environments with high levels of ambient noise, the noise picked up by the microphone will significantly impair the quality and/or the intelligibility of the transmitted speech signal. In order to get a reliable separation from the noise signal (e.g., engine noise, street noise), noise reduction algorithms have become part of digital speech coding systems recently. They are used, for example, in mobile communications, in hearing aids and in hands-free devices. Noise reduction systems always provide a tradeoff between speech quality and noise reduction. Hence, the aim is on the one hand to attenuate the noise signal n(k) in the output signal y(k) as much as possible and on the other hand to keep the distortion of the speech signal s(k) as low as possible at the same time.

Spectral weighting basically means that different spectral regions of the mixed signal of speech and noise are attenuated using different real valued factors between zero and one. The aim of this process is an audio signal which contains less noise than the original one. Besides requiring a minimal distortion of the original speech, it is also important that the residual noise, i.e., the noise remaining in the processed signal, does not sound unnatural.

Most state-of-the-art noise reduction systems are realized as depicted in Figure 2 and can be explained by means of noise estimation, spectral SNR estimation, and spectral weighting. In a first step the spectrum of the noise component is estimated given the noisy observation Y(λ, μ). Hereafter, the SNR is estimated exploiting the noise estimate and the last enhanced frame of the noise reduction system. If the SNR for a specific STFD coefficient is high an absolute gain close to one is chosen. In the opposed case where the SNR is low an absolute gain close to zero is applied. Thereafter, the processed spectrum is transformed back into the time domain.

The estimation of the short-term noise PSD remains a crucial and challenging task in every noise reduction system, especially in case of non-stationary noise. Several methods have been proposed for the estimation of the short-term noise PSD by tracking and post-processing the magnitude minima in the short-time Fourier domain, e. g., [heese15, gerkmann11, martin01, martin06].

Wind noise is a common type of background noise that is caused by air turbulences close to the microphone. Typical noise reduction techniques estimate the noise characteristics statistically, assuming the noise to be stationary. In contrast, wind noise is highly non-stationary due to a frequently changing flow velocity and chaotic turbulences. Therefore, specific estimators for wind noise are needed [nelke16].

First, the presence of wind noise has to be detected. At mobile phones, for example, this can be done by using two microphones: one on the front, the other on the back. Wind turbulences can be considered as a large quantity of independent sound sources for each microphone, whereas speech results from only one sound source. Therefore, the coherence between the two microphone signals can be exploited as a wind noise measure. Low coherence values indicate that wind noise is present [nelke13]. Another possibility is the usage of signal centroids [nelke14].

Secondly, the periodogram of the wind noise has to be estimated. The most difficult case is the concurrent presence of speech and wind. For vocals, the speech power is concentrated on harmonics, i.e. on multiples of the pitch frequency. Therefore, the noise power is estimated only at frequencies between the harmonics, using a pitch-adaptive binary mask, and interpolated otherwise [nelke15].

Single microphone noise reduction systems usually rely on different statistical properties of speech and noise. In addition, it is assumed that the ambient background noise is stationary or only slightly time-varying which is usually not fulfilled in practice. In consequence, statistical state-of-the-art noise estimators provide an estimate for the short-term noise power spectral density (PSD) in the best case. If the underlying noise signal exhibits a reasonable variance, the spectral fine-structure over frequency and time is estimated inadequately. Hence, statistical noise reduction systems are only able to remove the short-term mean of the noise which likely results in unpleasant artifacts that are called musical tones.

In contrast, the class of codebook based speech enhancement systems faces the aforementioned constraints by using a priori knowledge about speech and/or noise and also allows to model and thus cope with highly non-stationary noise environments [heese14; heese15a; heese16a]. Additionally, the codebook driven noise reduction systems have the potential to reduce the occurrence of musical tones, since the instantaneous speech and noise is estimated jointly over frequency and time.

With respect to the different advantages of statistical and codebook driven speech and noise estimators, an adaptive combination of different speech and noise estimates is desirable [heese16a]. In order to carry out this adaptive combination, a reliability measure is necessary. Utilizing the different speech and noise estimates, it is possible to compute all combinations of the speech and noise estimates which provide different estimates for the noisy observation. The distance between the different estimates and the noisy observation itself can serve as reliability feature. Closed form solutions can be derived for the optimal reliability weights and the resulting estimation error.

An example of Information Combining is depicted in Figure 4. A statistical based noise estimate and a codebook driven noise estimate yields after information combining the enhanced noise estimate.