Zoom out Search Issue

ManualsBrandsContents Manualsaudio & home theatreZoom in

IEEE SIGNAL PROCESSING MAGAZINE [28] MARCH 2015

signals y

and y

can be generated and presented to the left

and the right ear (cf. Figure 7).

In a bilateral system, i.e., a set of two independently operating

monaural systems, each device uses its own microphone signals

and optimizes its filter coefficients independently, which may

lead to a distortion of the binaural cues and hence the localiza-

tion ability [44]. To achieve true binaural processing, both

devices need to cooperate with each other and exchange informa-

tion or signals, e.g., through a wireless link. Currently, the first

commercial systems reach the market which exchange micro-

phone signals in full-duplex mode. These systems pave the way to

future implementations of fully

fledged binaural multimicrophone

signal extraction algorithms, where

microphone signals from both

devices are processed and combined

in each device. The gain in noise

reduction performance of a binaural

over a monaural system is exempla-

rily shown for an MVDR beamformer

in Figure 8.

The objective of a binaural speech enhancement algorithm is

not only to selectively extract the target source and to suppress

interfering sources and background noise, but also to preserve

the auditory perception of the complete acoustic scene. This can

be achieved by preserving the binaural cues, i.e., ITD, ILD, and IC,

of the target source and the residual interfering sources and back-

ground noise. In addition to monaural cues, these binaural cues

play a major role in spatial awareness and localization and are

very important for speech intelligibility due to binaural unmask-

ing, e.g., [5].

All discussed signal enhancement algorithms in the sections

“Data-Independent Beamforming” and “Statistically Optimum

Signal Extraction” essentially generate a single-channel output

signal. Since in a binaural system two output signals (i.e., one for

each ear) are required, this single-channel output signal can

either be binauralized, e.g., using binaural spectral postfiltering

techniques [19], [45], [46] or by mixing the output signal with

scaled (noisy) microphone signals [47], [48], or two different

complex-valued spatial filters can be optimized, where the desired

binaural cues are directly incorporated into the spatial filter

design, e.g., [48]–[50]. Although the latter paradigm allows for

more degrees of freedom to achieve noise reduction, there is typi-

cally a tradeoff between noise reduction performance and binau-

ral cue preservation.

In binaural spectral postfiltering

techniques, the same real-valued

gain is applied to one microphone

signal of each device, where a gain

close to one is applied when the

STFT bin should be retained (target

source), and a gain close to zero is

applied when the STFT bin should

be suppressed (interfering source or

background noise). This spectral gain can, e.g., be computed by

comparing the estimated binaural cues with the expected cues of

the target source or based on the temporal fluctuations of the

ITD [45]. Other commonly used approaches compute the spectral

gain based on the output signal of a data-independent or statisti-

cally optimum spatial filter (e.g., MVDR beamformer, BSS) [19],

[46]. Although binaural spectral postfiltering techniques preserve

the binaural cues of all sound sources, in essence, they can be

viewed as single-channel noise reduction techniques, hence typi-

cally introducing speech distortion and exhibiting single-channel

noise reduction artifacts (e.g., musical noise), especially at low

input SNRs.

The MVDR beamformer (using RTFs) and the MWF can be

straightforwardly extended into a binaural version producing

two output signals, by estimating the speech component in two

reference microphone signals, i.e., one on each hearing aid [48].

In [48] and [44], it was shown both analytically and using sub-

jective listening experiments that the binaural MWF preserves

the binaural cues of the target source but distorts the binaural

cues of interferers and noise, such that all components are per-

ceived as coming from the direction of the target source.

Clearly, this is undesired and, in some situations (e.g., traffic),

even dangerous. To optimally benefit from binaural unmasking

and to optimize the spatial awareness of the hearing aid user,

several extensions for the binaural MWF and the MVDR beam-

former have been proposed, which aim at also preserving the

binaural cues of the residual noise component by including cue

preservation terms in the binaural cost function, e.g., [48]–[50].

These include either RTF preservation or interference rejection

constraints for directional interfering sources [48], [49], or IC

preservation constraints for diffuse noise [50]. Another

approach is partial noise estimation, which corresponds to mix-

ing the binaural outputs with scaled versions of the noisy refer-

ence microphone signals [48].

[FIG8] The SNR gain of a monaural and a binaural MVDR

beamformer (diffuse noise field).

1,000

2,000

3,000

4,000

5,000

6,000

7,000

8,000

−2

SNR Gain (dB)

f (Hz)

Monaural (Two-Microphone) MVDR

Binaural (Four-Microphone) MVDR

TO EXPLOIT THE FULL POTENTIAL

OF BINAURAL PROCESSING,

BOTH DEVICES NEED TO

COOPERATE WITH EACH OTHER

AND EXCHANGE INFORMATION

OR SIGNALS.

THE WORLD’S NEWSSTAND

THE WORLD’S NEWSSTAND