We design and build wireless mic systems, mic-array devices, LE Audio and Auracast transmitters and on-device voice terminals. Clean capture, a link that holds up, consent built into the hardware, and audio that goes only to your cloud.
Speech models fail on bad input. A device built around your model gets cleaner audio than a phone or a laptop mic.
Concept render
Wireless mic
Two-person wireless mic kit
Two clip-on transmitters and one receiver for interviews, podcasts and field recording. Each transmitter keeps a local copy in case the link drops.
Concept render
Mic array
Mic-array meeting puck
A tabletop unit that picks up a whole room. Echo cancellation and beamforming decide whether the far end of the table gets transcribed.
Concept render
LE Audio module
LE Audio / Auracast USB transmitter
Plugs into a laptop, console or PA system and shows up as a standard USB audio device. Streams to earbuds, or broadcasts to many listeners with Auracast.
Concept render
Voice terminal
On-device voice terminal
A small speaker-and-mic device that hears its wake word locally and only then talks to your assistant. A mute switch and a light show when it listens.
System
From the air in the room to your model
An audio device is five layers. A mistake in the first one cannot be fixed in the last. We build the first four and test against the fifth.
01
Acoustics
Mic choice, port and gasket design, speaker chamber, and isolation between the speaker and the mics, tested in the real enclosure.
02
Signal processing
Echo cancellation, beamforming, noise suppression and gain control, tuned on recordings from your actual use case.
03
Codec and link
Opus, LC3, aptX, LDAC, AAC or uncompressed PCM, over 2.4 GHz, Bluetooth Classic, LE Audio, Wi-Fi or USB.
04
Device software
Firmware, signed OTA, pairing, settings and an SDK. App and backend in your cloud account are an option.
05 · YOURS
Your models
Your speech-to-text, voice assistant, translation or summaries. We deliver the audio your model expects and test against your endpoint.
Engineering record
A wireless recorder that never loses the take
Clip-on transmitters stream 2.4 GHz digital audio to a receiver and record to their own flash at the same time. If the wireless link drops, the local copy keeps the audio. Built by our Shenzhen team; client name withheld.
Audio
48 kHz / 24-bit WAV, uncompressed
Onboard storage
8 GB, up to 14 h, 30-min segments
When full
Loop overwrite of oldest segments
Link
2.4 GHz digital, one receiver, two transmitters
Range
Up to 250 m, line of sight
Runtime
7.5 h per transmitter, case recharge
Mass
33 g transmitter
Modes
Mono, stereo or safety track
Processing
Noise reduction, low-cut filter
Figures are from the product's specification sheet. Range depends on environment.
PCB layout
Wireless recorder transmitter, bottom layer.
PCB layout
Receiver, top layer.
PCB layout
Mic-array board, top layer.
Round mic-array board
A multi-microphone array for far-field voice pickup, the starting point for meeting pucks and voice terminals.
PCB layout
Strip recording card, receiver, top layer.
Strip recording card
A long, narrow recording card with matched transmitter and receiver boards.
Platforms our team builds on: Qualcomm QCC30xx and QCC51xx and JieLi audio SoCs for Bluetooth audio (LDAC, the aptX family, AAC, SBC, LC3, LE Audio and Auracast broadcast); a USB audio transmitter with codecs selected from an app; and an on-device voice assistant on Espressif ESP32-P4 with ESP32-C5.
Reference architecture
Inside a far-field voice device
A meeting puck and a voice terminal share the same chain. The front end and the wake word decide whether users have to repeat themselves.
Far-field pattern. Audio leaves the device only after the wake word or a button press, and only to your cloud.
Block
For example
Why it matters
Microphones
Knowles or TDK InvenSense digital MEMS mics, in a circular or linear array
Mic count and spacing set how well the beamformer separates a talker from the room
Front end
Acoustic echo cancellation, beamforming, noise suppression and automatic gain control, on a DSP core or the host SoC
Without echo cancellation the device cannot hear a user while its own speaker is playing
Wake word
A small keyword model running on the device
Audio stays on the device until the wake word or a button press. This keeps cloud cost and privacy exposure down
Host SoC
For example Espressif ESP32-P4 for audio and application processing, paired with an ESP32-C5 for Wi-Fi and Bluetooth LE
Sets the compute budget for the front end, the wake word model and any on-device speech features
Radio
Wi-Fi to your backend; Bluetooth LE for setup and phone pairing
A pre-certified module or a certified SoC reference layout keeps FCC work smaller
Cloud
Your speech or assistant service in your cloud account
By default we deliver the device protocol and SDK and test against your endpoint
Part names are examples only. Final selection happens at Gate 0 against your audio, latency, cost and module-origin requirements.
Key technologies
The engineering that decides whether the AI works
Most voice products that fail in the field fail on audio. The mic hears the speaker, the room echoes, or the link drops for a second at the wrong time. We work on these from the first prototype.
Acoustics in the enclosure. Mic port, gasket and speaker chamber design, measured in the real housing.
Voice front end. Acoustic echo cancellation, beamforming, noise suppression and automatic gain control, tuned on recordings from your use case.
Codecs. Opus, LC3, aptX, LDAC, AAC or PCM / WAV, chosen for what your model expects and what the link can carry. We measure the effect on your model's accuracy as well as bitrate.
Latency budgets. A budget per stage (capture, processing, encode, radio, decode) and an end-to-end measurement on the prototype.
Robust 2.4 GHz links. Interference handling in crowded rooms, and a local backup recording on the transmitter so a dropout does not lose audio.
LE Audio and Auracast. LC3 streams to earbuds and hearing devices, and Auracast broadcast to many listeners at once.
USB Audio Class. Transmitters and dongles that work as standard USB audio devices, without a driver install.
Wake word on device. A local keyword model, so the device does not stream audio until someone asks it to.
Certification path
What an audio device must pass before you sell it in the US
Most audio products here are FCC Part 15 devices. Adding Bluetooth adds SIG qualification. We map yours at Gate 0 and file in your company's name.
Requirement
When it applies
How we handle it
FCC Part 15 Subpart C (§15.247)
Bluetooth, Wi-Fi and most 2.4 GHz digital audio links
Covers frequency hopping and digitally modulated transmitters in 2400–2483.5 MHz. Use a pre-certified module where the form factor allows. A chip-down design needs its own intentional-radiator testing.
FCC Part 15 Subpart C (§15.249)
Some low-power 2.4 GHz proprietary links
An alternative section with a field-strength limit of 50 mV/m on the fundamental. The lab confirms whether your link is tested to §15.247 or §15.249.
FCC Part 15 Subpart B
Every device with digital circuitry
Unintentional emissions. Still required when the radio module is pre-certified. Receivers and USB transmitters need it too.
RF exposure (47 CFR §1.1307, §2.1093)
Body-worn transmitters and anything used within 20 cm of the body
A clip-on transmitter is a portable device. Sources at or below 1 mW time-averaged power are exempt from evaluation. Above that, the lab applies the exemption formulas or runs SAR tests.
Bluetooth SIG qualification
Any product that uses Bluetooth
Every Bluetooth product must complete the Bluetooth Qualification Process before it is sold, under your company's SIG membership. A module supplier cannot qualify it for you. A product qualification fee applies to each new design.
UN38.3 + 49 CFR §173.185
Any lithium cell shipped with or in the device
Each cell or battery must be of a type proven to meet UN Manual of Tests and Criteria, Part III, sub-section 38.3. Plan it with the cell supplier at Gate 2.
California Penal Code §632 (not a certification)
Any product that records conversations
California requires the consent of all parties to record a confidential communication. We design for visible, provable consent: a recording light and physical controls.
How a 10–1,000 units run works for an audio device
Audio products are judged by ear. Each gate puts working units in your team's hands early, so you hear the result before the enclosure is locked.
Gate 1, works-like (3–5 units, 3–6 weeks). Proven audio SoC and mic boards on a custom carrier. Your team records real sessions and runs them through your model.
Gate 2, EVT (10–30 units, 6–10 weeks). First custom PCB and enclosure. We measure frequency response, echo performance and link range in the housing.
Gate 3, pilot (10–1,000 units, 8–16 weeks). A pilot with real users. Certification and Bluetooth qualification run in parallel.
Every unit off the line goes through a production test we write with you:
Mic sensitivity and frequency response in an acoustic test box, against limits set from EVT data
Speaker frequency response and distortion check, where the device has a speaker
RF conducted power check on the 2.4 GHz or Bluetooth radio
Pairing test between transmitter and receiver, or with a reference phone
Codec check: a known test signal through the full encode and decode path
Recording light, button and charge check; battery capacity sampled per lot
No Duobay server and no China-hosted service in the audio path. Devices talk to your backend or to a US cloud region you choose.
The wake word runs on the device. Nothing is sent until the wake word or a button press.
A hardware mute switch and a recording light the user can see. Firmware cannot hide when the mics are live.
You generate and hold the signing keys and device credentials. We can run key injection with keys your team provides.
You receive the firmware source and a software bill of materials (SBOM) listing every library, version and license, including the audio DSP and codec stacks.
Audio and voice hardware: common questions
Should our wireless mic use Bluetooth or a proprietary 2.4 GHz link?
It depends on what the receiver is. If the audio goes to a phone or earbuds, Bluetooth is the practical choice. If you control both ends and need uncompressed audio, longer range or two transmitters into one receiver, a proprietary 2.4 GHz digital link usually works better. We compare both at Gate 0 against your range, latency and battery targets.
We use a qualified Bluetooth module. Do we still need Bluetooth SIG qualification?
Yes. As of October 2026 the Bluetooth SIG states that every Bluetooth product must complete the Bluetooth Qualification Process under the selling company's own membership, and that suppliers cannot qualify products on your behalf (bluetooth.com). Using a qualified design keeps the work smaller. We prepare the submission with you.
How do you handle recording consent?
California requires the consent of all parties to record a confidential communication (Penal Code §632, as of October 2026). Our default is a visible recording light and a physical record or mute control. We do not design covert or hidden recording products. Your counsel decides how consent works in your markets.
Do you build licensed UHF wireless microphones?
No. Licensed wireless microphones operate as low power auxiliary stations under FCC Part 74, Subpart H, and unlicensed TV-band microphones fall under §15.236. Those are out of scope for us. We build in the 2.4 GHz band, with Bluetooth, Wi-Fi and proprietary digital links.
Can a voice device work without sending audio to the cloud all the time?
Yes. The wake word runs on the device, so audio leaves only after the wake word or a button press. If your model is small enough, some commands can run fully on the device. We size the SoC for that at Gate 0.
Can we start from boards you have already built?
Often, yes. Our team has built a wireless recorder system and mic-array boards, and the voice-capture badge reference design has the mic, codec and power work done. Starting from proven boards can shorten Gate 1 and Gate 2. Your design and source files are still yours.
Next step. Tell us what your device needs to hear, where the audio goes and how many units you want for a first pilot. A Duobay engineer replies within 2 business days.
Concept render
PCB layout
Concept render
Concept render
Concept render
PCB layout
PCB layout
PCB layout
Illustration
Illustration