Mảng chùm tia và micrô

A microphone array is a set of two or more microphones mounted at fixed, known positions, for example a ring of seven on top of a smart speaker. Beamforming is the signal processing that combines their outputs into one channel that listens mainly in a chosen direction, the look direction. It works like a spotlight for sound: speech from the target direction is reinforced while noise, music and other voices from elsewhere are reduced. Nothing physically moves; the beam is steered entirely in software, which is why one device can follow a person walking around a room.

The simplest method is delay-and-sum. Sound travels at about 343 meters per second, so a voice arriving from the left reaches the leftmost microphone slightly before the rightmost one. The delay between two microphones equals their spacing times the cosine of the arrival angle, divided by the speed of sound; for microphones 5 centimeters apart and sound arriving end-on, that is about 146 microseconds. Delay-and-sum shifts each channel by the delay expected for the look direction, so the target lines up perfectly, then adds the channels. The aligned target adds coherently, while noise that differs at each microphone adds up only partially. MathWorks' worked example shows the payoff: with 10 elements, the array gain is 10, so output signal-to-noise ratio is about ten times that of a single sensor for uncorrelated noise.

In practice audio beamformers work in the frequency domain. The signal is cut into short frames and converted with a Fourier transform into frequency bins, and each bin gets its own complex weight per microphone, which applies a phase shift and gain. This matters because a fixed time delay corresponds to a different phase at every frequency. Geometry sets hard limits. If microphones are more than half a wavelength apart, the array cannot tell some directions apart and produces ghost beams, an effect called spatial aliasing; at 8 kHz the wavelength is about 4.3 centimeters, so spacing must stay near 2 centimeters to avoid it. At low frequencies, where wavelengths are meters long, a small array barely differs from one microphone, so its beam is very wide.

Adaptive beamformers go further by also learning where the noise is. The minimum variance distortionless response (MVDR) beamformer chooses weights that keep the look direction at exactly unit gain, the distortionless constraint, while making total output power as small as possible. Since the target must pass unchanged, the only way to lower output power is to suppress everything else, so MVDR automatically places deep nulls on strong interferers. In MathWorks' example it put deep nulls on interferers at 30 and 50 degrees while keeping 0 dB gain toward a target at 45 degrees, where plain delay-and-sum let the interference mask the signal. The linearly constrained minimum variance (LCMV) beamformer generalizes this with several constraints at once.

Deep learning changed how the look direction is found. In 2016 Heymann, Drude and Haeb-Umbach at Paderborn University, and separately Erdogan and colleagues at MERL, trained neural networks to estimate a mask, a value for every time-frequency point saying how likely it is to be speech rather than noise. The masks are used to estimate the spatial statistics of speech and of noise, which in turn give MVDR weights without knowing the talker's angle in advance. Google described neural networks that perform the multichannel processing jointly with acoustic modeling in its 2017 paper on Google Home, reporting 8 to 28% fewer word errors relative to the existing production system. Apple's December 2018 article on HomePod described a six-microphone chain of echo cancellation, mask-based multichannel Wiener filtering, dereverberation and blind source separation, with the "Hey Siri" detector choosing the best output stream. Apple reported word error rate improvements of about 40% for reverberant speech, 74% in loud background noise and 61% with a competing talker.

Compared with single-microphone noise suppression, beamforming has an advantage that no single-channel model can match: it uses where the sound comes from, not just what it sounds like. A linear beamformer such as MVDR also adds little distortion to the target, which is why it is popular in front of speech recognizers. The trade-off is hardware: more microphones, synchronized sampling and careful placement. Hearing aids show the scale of benefit with only two microphones per ear. Laurel Christensen's 2013 review in The Hearing Review puts first-order directional benefit at about 3 dB in real life, notes that linking the microphones on both ears narrows the beam from about 60 to about 45 degrees, and reports that this added about 1 dB in speech-shaped noise but nothing measurable in babble.

Beamforming has well-known failure modes. Reverberation is the biggest: in a room, a voice reaches the array directly and also via many reflections from every wall, and reflected energy arriving from all sides cannot be removed by aiming. The Hearing Review notes that directional hearing aids help mainly when the talker is within about 10 feet, inside the critical distance where direct sound dominates. Diffuse noise, such as a crowded restaurant, has no single direction to null, so adaptive hearing aids fall back to a fixed pattern. Steering errors are dangerous for MVDR: in MathWorks' example, aiming at 43 degrees when the talker was at 45 made MVDR treat the talker as interference and suppress it, which is called signal self-nulling. Moving talkers, microphone mismatch and loudspeaker echo add further error.

What it takes in practice depends on the product. A hearing aid manages with two microphones a few millimeters apart and tight power limits. A smart speaker uses six or seven microphones in a ring of a few centimeters, multichannel echo cancellation so it can hear while playing music, and enough compute to run continuously; Apple says its HomePod front end uses less than 15% of one core of a 1.4 GHz A8 chip. MVDR itself requires estimating and inverting a small covariance matrix, one row and column per microphone, for every frequency bin, a modest cost on modern chips. The harder engineering is mechanical: keeping microphones matched, sealed from internal vibration and placed according to the spacing rules.