Quectel Opens Circular Microphone Array Solution for Evaluation: In-House Edge AI Audio Algorithms with the VA60X-GL Six-Mic Ring Board

On 29 October Quectel released a circular microphone array solution and opened it to all customers for evaluation. The package pairs the company's in-house edge AI audio algorithms, running on the board, with VA60X-GL, a trimmable ring-shaped capture board. Acquisition, AI noise reduction, sound-source localisation, wake-up and response all happen at the edge. The target applications named in the announcement are robots, smart speakers and conference-room microphones.

The interesting part is not another audio board. It is where the delivery boundary now sits. Building far-field voice into a product used to mean a bespoke acoustic front end: microphone selection, array geometry, AEC and beamforming tuning, each sourced separately, with integration cycles measured in months. Bundling algorithm, capture board and a standard USB audio path turns that front end into a purchasable part. Evaluation moves from "define the architecture, then build a prototype" to "test the board, then define the architecture".

What is actually shipped: algorithms on the board, UAC at the interface

VA60X-GL is a ring capture board with six high-SNR analogue microphones in its standard configuration. Audio leaves over USB Type-C under the UAC (USB Audio Class) standard, so it is driver-free and plug-and-play. The technical consequence is that the host needs no audio driver: Android and Linux see a standard USB microphone, and the signal has already been processed on the board. I2S, I2C, UART and additional microphone interfaces are reserved for connecting mainstream Android or Linux host modules, and for the customer's own speaker and reference-signal wiring.

Placed next to Quectel's existing audio products, the positioning is distinct. The previously published VA500-GL is an audio module: 16 kHz sampling, offline wake word and command recognition, vendor-quoted recognition above 95% and response under 100 ms, 92% command accuracy at 65 dB background noise, 5V supply, up to 2.85W output, two UARTs, -40 to 85°C, 23mm × 20mm × 3.3mm, aimed at air conditioners, purifiers and range hoods. VA60X-GL takes the other route: not single-point near-field commands but a 360° far-field front end, with the algorithm weight on beamforming, noise reduction, dereverberation and echo cancellation. One is about hearing a sentence; the other about continuously hearing who is speaking in a noisy room.

The checkable specifications

Six high-SNR analogue microphones in the standard version, trimmable to three or four for specific use cases;
Ring diameter customisable against the customer's mechanical design (no standard diameter is published);
UAC over USB Type-C, driver-free and plug-and-play;
I2S, I2C, UART and microphone expansion, connecting to mainstream Android and Linux host modules;
Not disclosed: microphone SNR and unit-to-unit consistency, board power, algorithm compute load, standard diameter, pricing, MOQ and lead time.

The last point is not nitpicking — it is the first thing a selection engineer needs. In any array, the performance ceiling is set by microphone consistency: sensitivity and phase spread within a batch determine how far down the beamformer can push side lobes. The announcement gives only "high SNR" for this. Without the numbers, an evaluation can reach qualitative conclusions only.

Four algorithm claims and how well each is evidenced

Four capabilities are listed, and they do not carry equal evidential weight.

Beamforming and localisation: multiple microphones form beams in different directions, locating the source across 360°, extending capture distance and suppressing side and rear interference. Described qualitatively, with no angular error, no range figure and no recognition-versus-distance curve.
Deep noise reduction and dereverberation: a neural network handles stationary noise (air conditioning, fans), non-stationary noise (keyboard clicks) and room reverberation. No before-and-after SNR improvement is given.
Echo cancellation: test data cited for the extreme case of a microphone mounted directly against the loudspeaker playing at 100 dB, where wake-up and interruption still work. This is the only claim in the release carrying both a condition and a number.
AI vision-assisted tracking: combined with Quectel's vision algorithms, the system keeps tracking a speaker through short silences, so a servo or gimbal turns smoothly without waiting for the next command. Quectel says this is already deployed on its in-house Mamboe robot.

The 100 dB figure needs context. That is roughly the sound pressure of a subway train arriving or a small live show, and a microphone pressed against the loudspeaker means an extremely short echo path with direct sound far above reverberation — the regime where linear coupling is strongest, non-linear distortion is most concentrated, and AEC most often breaks down. Holding wake-up and interruption under those conditions says the on-board echo path and residual suppression are tightly engineered.

But the measurement establishes robustness at a limit, not usability in a product. A robot is a metal or plastic enclosure with structure-borne conduction, multi-path reflections and its own joint noise; the AEC reference signal passes through an amplifier and a cavity that both add non-linearity. None of that appears on a bench. The release gives no test room, no echo-path length, no ERLE and no double-talk behaviour, so 100 dB is evidence that a capability exists, not a specification that transfers to a finished machine.

How it compares with published figures elsewhere

A six-microphone ring is not a new architecture. AISpeech publishes a circular six-mic development kit on its DUI platform with 360° pickup, ±10° source localisation, 92% far-field recognition at five metres, 95% wake-up rate and a 72mm board diameter. That kit has been on sale for years, which is enough to establish that ring array, beamforming and wake-up are mature building blocks.

The two vendors measure different things. AISpeech publishes performance scales — angular error, recognition against distance. Quectel publishes an engineering boundary — behaviour under extreme echo. The two cannot be compared directly, and neither establishes superiority. What is comparable is the delivery condition: three- or four-microphone trimming, a customisable diameter, driver-free UAC. For a customer those are the variables that set the length of an evaluation: whether the board fits an existing enclosure and runs on an existing host decides whether the project proceeds at all, earlier than one degree of localisation accuracy does.

Why now

There is a clear demand backdrop. IDC puts global humanoid robot shipments at close to 25,000 units in the first half of 2026, up 432.1% year on year; China accounted for more than 19,000 units, up 426.3% and about 77.9% of the total, and IDC has raised its 2030 forecast to more than 750,000 units. What robots need from audio is not a single command heard clearly but 360° discrimination of who is speaking, wake-up that survives the machine's own motion noise, and interruption at any moment — exactly the case for a ring array rather than a single microphone or a linear array.

The device side grows far more slowly. GII's High SNR MEMS Microphones report puts the global market at USD 2.63 billion in 2026, reaching USD 3.89 billion by 2032, a 6.69% CAGR. Microphones are not a high-growth category; the added value sits in the algorithm and the assembled board. That explains why a module vendor would rather sell the board than the parts.

Supply-side motivation is equally visible. Counterpoint Research recorded 11% year-on-year growth in global cellular IoT module shipments in Q2 2026, with Quectel's share rising to 41% from 35% a year earlier. Its customer base already contains a large number of Android and Linux platform projects, and an audio capture board is an adjacent part sold into the same accounts, the same mechanical engineers and the same certification flows. For the vendor it is incremental revenue on existing channels; for the customer it merges two suppliers into one.

HSY Perspective

What changes here is not pickup accuracy but how an acoustic front end is procured: from a bespoke algorithm delivery to a board that can be tested the moment it is plugged in. The six-mic ring is mature architecture — AISpeech and others have published 360° pickup, ±10° localisation and 92% recognition at five metres for years. Quectel offers no new number on angular error or recognition rate; what it publishes is a delivery condition: three- or four-mic trimming, a customisable diameter, driver-free UAC, reserved I2S and UART. Those describe friction in evaluation and ramp, not a performance ceiling.

Both sides of the market support the reading. On demand, IDC counts close to 25,000 humanoid units globally in the first half of 2026 with China at 77.9% — machines whose audio requirement is 360° discrimination and continuous wake-up. On supply, Counterpoint puts Quectel's module share at 41% in Q2 2026, into a base already running Android and Linux hosts, where an audio board rides the same accounts and the same mechanical validation. Mature architecture plus an existing channel makes a purchasable front end viable.

Two boundaries apply. If UAC carries only pre-processed audio, swapping the algorithm means rebuilding outside the board and the front end is locked to the supplier; the release does not say whether the output is processed mono or raw multi-channel, and that detail decides whether this is a replaceable component or a tied solution. Fewer microphones also weaken spatial resolution and side-lobe rejection directly, and a three-mic ring discriminates front from back less well than a six-mic one, so the saving returns as a penalty in difficult acoustic fields. The observation point is concrete: whether the board reaches a published price list and a lead time quoted in weeks.

Sources: Quectel press release, "Edge AI reshapes hearing: how Quectel gives intelligent terminals a 360° ear", 29 October 2026, quectel.com.cn; Quectel AI audio module solution page (VA500-GL specifications), quectel.com.cn; IDC Worldwide Humanoid Robot Tracker, H1 2026; Counterpoint Research Global Cellular IoT Module and Chipset Tracker, Q2 2026; GII, High SNR MEMS Microphones Market, Global Forecast 2026-2032; AISpeech DUI circular six-microphone development kit specifications.