Abstract: AI glasses, meeting recording devices, in-vehicle cabins, and robots are pushing voice from a "wake-up entry" to a "productivity entry." Yan Gengzhen, Director of Soundskrit China, believes that although the traditional "omnidirectional microphone + array + algorithm" solution is mature, it faces challenges in miniaturized terminals such as directional performance loss, microphone spacing, algorithm complexity, and mass production consistency. Soundskrit's approach is to shift directional information forward to the physical layer of MEMS microphones. By outputting raw signals with spatial directivity through directional microphones and combining them with algorithms, it provides AI terminals with more reliable target sound source enhancement capabilities.
Over the past two years, the hype around AI hardware has been almost entirely dominated by models, computing power, and form factor innovations. AI glasses, AI toys, recording cards... various form factors have emerged densely, and terminal manufacturers are all trying to answer the same question: How can large models move from the cloud to the real world?
However, in real-world scenarios, a more fundamental issue is beginning to surface—AI devices must first hear clearly what people are saying. In the past, voice interaction mostly stayed at the level of wake words, simple commands, and smart speaker Q&A. Occasional false wake-ups or recognition failures were still acceptable to users. But as voice begins to enter scenarios such as meeting minutes, interview transcription, live broadcast audio capture, multi-speaker attribution, voiceprint recognition, and even "anytime, anywhere personal assistants," the quality of audio input is no longer just an experience issue; it directly determines whether the results are reliable.
"I feel things are different now. Audio is now used for business purposes, and the requirements are somewhat higher," said Yan Gengzhen, Director of Soundskrit China, in an exclusive interview with EEFOCUS during the recently concluded electronica Munich, where he also conducted product demonstrations and explained the technical principles on-site.
Why Are Traditional Omnidirectional Microphone Arrays Increasingly Struggling in AI Terminals?
In today's mainstream audio solutions, the "omnidirectional microphone + array + algorithm" approach remains the most common and mature route. Its implementation principle is as follows: omnidirectional microphones first capture surrounding sounds indiscriminately, and multiple microphones form an array. The direction of the sound source is determined through the time difference, phase difference, and amplitude difference between different microphones. Then, algorithms perform processing such as beamforming, differencing, noise reduction, and echo cancellation, ultimately "extracting" the target sound from complex environmental noise.
Yan Gengzhen does not deny the maturity of the traditional array route, but he believes this route is approaching a technical bottleneck in AI terminals: "The more you want to make it usable in complex environments, the 'heavier' the system becomes." In his explanation, the problem with traditional arrays is first reflected in "effective performance loss." A single omnidirectional microphone itself may have a good signal-to-noise ratio (SNR), but it lacks natural directivity. To achieve directional sound pickup, it is necessary to use two or even more omnidirectional microphones for differencing, constructing a "figure-8" like beam through algorithms. This process does not come without a cost. Yan Gengzhen cited common industry specifications as an example: a single omnidirectional microphone might achieve about 70dB, but after differencing two omnidirectional microphones to achieve directivity, the system's effective performance might drop to the 51-55dB range.
The directivity of traditional arrays does not come from the microphone device itself, but from backend computation. Sufficiently obvious spatial differences must first be captured between two omnidirectional microphones before the algorithm has the conditions to determine which direction is the target sound source and which is noise. However, differencing itself cancels out part of the signal and also raises the system's noise floor. In traditional audio application scenarios, this approach is sufficient, but AI terminals face more complex application scenarios.
Figure-8 Beam Pickup Area
"This is a simple diagram of our single-chip directional microphone: the single chip does not need algorithms; raw data output can form a good figure-8 beam pickup area, isolating noise from the 90-degree and270-degree directions." Traditional omnidirectional microphone arrays rely on algorithms to "calculate" directivity at the back end, whereas directional microphones aim to have directional information at the raw signal stage.
The second bottleneck comes from size. For a microphone array to function effectively, a certain distance must be maintained between the microphones because the algorithm relies on the differences in sound waves captured at different positions. Yan Gengzhen relayed industry experience, stating that when using two omnidirectional microphones for directivity, the spacing usually needs to reach about 2 centimeters to be more reliable. This distance is not exaggerated for desktop conference equipment or smart speakers, but when placed in AI devices, the design size becomes too large. If directional pickup must rely on an array spacing of several centimeters, many product form factors will naturally be constrained.
The third bottleneck is system complexity. Traditional arrays can solve problems with two microphones in simple scenarios, but once entering scenarios such as 360° conference rooms, multi-speaker attribution, four-seat voice zoning in vehicles, and robot sound source localization, manufacturers often need to increase the number of microphones, using four-microphone, six-microphone, or even larger arrays to improve spatial sampling capabilities. As the number of microphones increases, problems also multiply: structural openings become more complex, acoustic consistency is harder to guarantee, the algorithm tuning cycle lengthens, and the pressure on power consumption, cost, testing, and mass production calibration all rise. On some recording card devices, traditional solutions might use 5 or 6 microphones for an array, whereas Soundskrit aims to achieve a similar effect with a compact combination of "two directional microphones + one omnidirectional microphone."
Heavier algorithms are not necessarily better. Many AI hardware devices need to reduce environmental noise while preserving the naturalness of the human voice. Yan Gengzhen mentioned that some scenarios do not want "heavy processing," especially those aiming to preserve the original texture of speech and music. Excessive noise reduction will bring distortion and a sense of unnaturalness, ultimately returning to the problem of "the input end not being clean enough."
In fact, the audio requirements for AI devices are no longer just simple sound capture, but knowing who is speaking and from which direction. This is also the reason Yan Gengzhen repeatedly emphasizes "capturing the sound cleanly first." If too much environmental noise, multiple voices, and spatial reflections are mixed into the input end, the backend model will have to pay for heavier computation, more aggressive filtering, and more complex compensation. For transcription, speaker diarization, speaker attribution, and voiceprint recognition, the front-end signal quality itself is a fundamental condition.
What AI terminals need is a more fundamental "enhanced audio" capability: separating the target sound source from environmental noise at the front end as much as possible, and bringing spatial direction, sound source location, and target user information into the signal chain earlier.
Shifting Directivity Forward to the Physical Layer of Microphones
If the idea behind traditional omnidirectional microphone arrays is to "first capture the sound, and then separate the target sound through algorithms," what Soundskrit wants to do is exactly the opposite: enabling the microphone to have directivity at the raw signal stage.
According to Yan Gengzhen's description, Soundskrit's solution still belongs to the MEMS microphone route. Soundskrit's directional microphone has a hole opened on both the top and bottom of the vibrating membrane, allowing sound to enter the membrane from different directions. This introduces usable path differences, phase differences, and pressure differences, helping the system identify the direction of the main speaker and suppress lateral noise at the physical layer. Yan Gengzhen summarized it in one sentence: "Our silicon microphone has a hole opened on both the top and bottom of the vibrating membrane... identifying who the main speaker is through phase?"
This principle can be understood in a more intuitive way. The goal of an ordinary omnidirectional microphone is to capture sound from all directions as uniformly as possible; whereas Soundskrit's directional MEMS microphone is more like a "direction selector" pre-designed in the device structure. When sound comes from directly in front of or directly behind the microphone, there are differences in the path, time, and pressure of it reaching both sides of the membrane. The membrane will vibrate accordingly and convert it into an electrical signal. However, if the sound comes from a lateral direction of 90 degrees or 270 degrees, ideally, the time and attenuation distance for it to reach both sides of the membrane are closer, the pressure difference on both sides of the membrane becomes smaller, and the vibration weakens accordingly. Yan Gengzhen explained on-site that if the pressure of lateral noise reaching both sides of the membrane is close, "then it won't vibrate... and won't generate an electrical signal." This means the microphone has naturally suppressed lateral noise at the physical layer.
The result this brings is that the raw output of a single directional microphone can form a "figure-8" like pickup area. Yan Gengzhen held the physical product on-site to explain that through the structural design of the two holes, the Soundskrit microphone extracts the strongest sound in the 0-degree and 180-degree directions, while the sound in the 90-degree and 270-degree directions is significantly suppressed.
This is exactly the meaning of "shifting directivity forward to the physical layer."
Demonstration Prototype for 3D Spatial Sound Pickup
Soundskrit's Butterfly evaluation kit further demonstrates the extension of this approach in 3D spatial sound pickup. Butterfly includes three orthogonally arranged Soundskrit dipole microphones and one omnidirectional microphone, with a total thickness of less than 3.5mm. The microphone signals can be used to determine the direction of sound arrival and automatically steer the directional beam towards the sound source; its 3D auto-steering mode can detect the loudest sound source in any direction and display the azimuth and elevation angles.
In Soundskrit's product system, Butterfly is more like a demonstration prototype that "visualizes" the underlying technology. It allows terminal manufacturers to intuitively see where the sound comes from, how the system determines the direction, and how it steers the pickup beam.
Structurally, the key to Butterfly lies in the establishment of "spatial coordinates." The X/Y two dipole microphones cover the horizontal plane, the Z-axis dipole microphone covers the vertical direction, and the W omnidirectional microphone provides a reference signal. In this way, what the system processes is no longer a set of ordinary audio signals, but sound input with spatial directional information. For scenarios such as meeting transcription, smart cabins, and robot interaction, the significance of this capability is very direct: the device needs to know whether the speaker is coming from the left, right, front, or from a higher or lower position.
During the on-site demonstration, Yan Gengzhen focused on showcasing the effect of this "direction recognition + beam tracking." He had on-site personnel speak, and the interface displayed the planar angle; when the person turned or moved to different directions, the system outputted the angle changes accordingly. He mentioned that at a distance of one meter, planar angle recognition can achieve about ±2.5 degrees; after switching to 3D mode, it can also recognize the up and down direction of the sound, that is, adding vertical angle judgment in addition to the horizontal angle. According to his explanation, such capabilities can correspond to applications such as smart cabins, robot prototypes, and meeting transcription. "It actually forms a pickup area in this direction while recognizing this angle," thereby instantly suppressing noise from other directions.
Therefore, the value of Butterfly lies in achieving 3D sound source localization and automatic beam steering with an extremely small volume. For customers, such prototypes can turn "spatial audio input" from a concept into a testable, debuggable, and transferable design reference.
From AI Glasses to Recording Cards, Directional Sound Pickup Is Entering Real Product Form Factors
The value of directional sound pickup must ultimately be verified in specific product form factors. At the electronica Munich, Yan Gengzhen first showcased a pair of AI glasses. Its special feature is the placement of a directional microphone near the bridge of the nose, making the pickup direction closer to the wearer's mouth.
When users make phone calls, record audio, or interact with AI through the glasses, the system prioritizes extracting sound near the mouth, avoiding capturing all surrounding sounds at once.
During the on-site test, the author kept speaking after answering a call, while Yan Gengzhen stood nearby creating interference sounds, simulating close-range noise at the exhibition site from different positions on the left, right, and below. Ordinary omnidirectional microphones easily capture nearby human voices into the call link, but in this demonstration, the nearby noise was significantly suppressed, and the pickup area was more concentrated on the wearer's mouth.
Glasses are naturally close to the user and also naturally exposed to open environments. Traditional omnidirectional microphones on them easily capture passersby's voices, wind noise, and environmental noise together; the significance of directional microphones lies in first defining the effective pickup area, allowing the backend algorithms to face clearer input.
Another category of AI products that is rapidly heating up is the recording card. Compared to traditional voice recorders, these devices are lighter, thinner, and place more emphasis on the AI workflow of "organize right after recording." Yan Gengzhen mentioned that on some recording card devices, traditional solutions might require 5 or 6 microphones for an array, whereas Soundskrit hopes to achieve a better spatial pickup effect through a compact combination of "two directional microphones + one omnidirectional microphone."
This approach is also extending to in-vehicle cabins and robots. Yan Gengzhen mentioned that smart cabins will increasingly focus on "voice zoning" in the future, including front and rear row isolation and identifying which seat is speaking; robots, on the other hand, need to determine the user's direction in noisy environments before responding. For these products, directional sound pickup has evolved from simple audio quality optimization to a part of AI terminals understanding real-world scenarios.
Conclusion: The Audio Competition in AI Hardware Has Just Begun
Soundskrit is a directional microphone technology company from Quebec, Canada. Its core direction is to shift the directional information of sound forward to the MEMS microphone device level. Through directional silicon microphones, DOA (Direction of Arrival) sound source direction recognition, and beamforming capabilities, it provides enhanced audio solutions for scenarios such as AI glasses, recording cards, meeting transcription devices, smart cabins, robots, earphones, and industrial noise reduction. Its technical feature lies in the ability of a single microphone to form directional pickup capability. Coupled with multiple directional and omnidirectional microphones, it can also achieve 360° sound source tracking and 3D spatial sound pickup.
From AI glasses and recording cards to in-vehicle cabins and robots, the core value of future audio systems will increasingly approach "perception." It needs to help devices determine who is speaking, where they are speaking from, and whether it is the target user speaking. The ultimate competition in AI hardware is not just about the ability to answer, but also the ability to obtain reliable input in real-world environments. Only when devices can first hear clearly, distinguish clearly, and match correctly, can the backend intelligence truly have room to play.
The inspiration from Soundskrit's solution lies in the changing division of labor in the audio chain. The hardware front-end first completes part of the spatial selection and target sound source enhancement, and then the algorithm takes charge of more complex sound source localization, beam switching, voiceprint recognition, multi-speaker diarization, and semantic understanding. The 3D sound source direction recognition and automatic beam steering capabilities demonstrated by the Butterfly evaluation kit also indicate that spatial audio input is becoming a callable front-end capability for AI terminals, rather than just an add-on for recording quality.
Yan Gengzhen emphasized in the interview that if the front-end cannot capture the target sound cleanly, the backend ASR (Automatic Speech Recognition), voiceprint recognition, and large models will all be affected. Therefore, the audio competition for AI terminals will increasingly focus on "physical layer enhancement + algorithm collaboration," and this competition has just begun.