What makes an AI voice device hard to manufacture
An AI voice device is a microphone array, an audio front end that cancels echo and suppresses noise, a wake-word engine, a processor with Wi-Fi, a speaker and amplifier, and an enclosure that works as part of the acoustic system. The language model usually runs in the cloud. The manufacturing challenge is making far-field voice pickup work while the device’s own speaker is playing, in every unit, which puts the enclosure, the microphone sealing, the speaker and production acoustic testing at the center of the program.
Teams coming from software often prove the experience on a development kit with a USB microphone array, then find that the product enclosure changes almost everything: the echo path, resonances, rattles and the spread of microphone sensitivity between units. Plan acoustic design and acoustic test fixtures as early as the industrial design, and decide early how much of the AI runs on the device, because that choice sets the processor, memory, thermal design and cost.
Where the intelligence runs
| Architecture | Hardware | Trade-offs |
|---|---|---|
| Cloud-first | Wi-Fi microcontroller or low-cost SoC; wake word on the device, audio streamed after wake | Lowest bill of materials; latency and quality depend on the network; ongoing cloud and model cost per active user |
| Hybrid | Application processor with NPU or DSP; wake word, voice activity detection and some speech recognition or a small model on the device | Faster and partly offline; more memory, heat and cost |
| Phone-tethered | Bluetooth Low Energy device with its own battery, using the phone for connectivity and compute | Small and portable; depends on the companion app and on the phone’s background limits |
Cloud cost does not appear on the bill of materials, but it often decides whether the business works. Model it per active user per month before you fix a hardware price.
Key components and sourcing
- MEMS microphones. Digital or analog, chosen for signal-to-noise ratio, acoustic overload point and sensitivity tolerance. Beamforming depends on matched microphones, so tighter-tolerance parts or per-unit calibration are worth paying for.
- Audio front end. Echo cancellation, beamforming and noise suppression, running on the main processor, a dedicated audio DSP or licensed software. Put the licensing terms and the tuning responsibility in writing.
- Wake-word engine. Licensed or built in-house. Validate it on audio recorded through your own enclosure, not only on the vendor’s data.
- Speaker and amplifier. Driver choice, enclosure volume and a sealed, ported or passive-radiator design set bass response and distortion. Drivers vary from unit to unit, so set incoming limits on sensitivity and resonant frequency.
- Processor and wireless. Pre-certified Wi-Fi and Bluetooth modules reduce radio testing.
- Privacy hardware. A mute switch that physically cuts the microphones, and an indicator that software alone cannot turn off.
Acoustic decisions to make before tooling
- Microphone placement and spacing. The array geometry has to suit the beamforming algorithm. Keep the microphones as far from the speaker as the form allows, and isolate them from enclosure vibration.
- Microphone ports. Each port needs a short, sealed path from outside to the microphone. A leaking gasket lets in sound from inside the enclosure and changes sensitivity and directivity. Add mesh or a membrane against dust and liquids, and measure its acoustic effect.
- Speaker enclosure. Volume, sealing and bracing decide bass and distortion; screw torque, foam and cable routing decide whether the product rattles.
- Distortion and echo cancellation. Echo cancellers handle linear echo well. Speaker distortion and rattles create nonlinear echo that they handle poorly, so users end up shouting over their own music. Limiting maximum volume or adding dynamic range control is often better than a louder product that cannot hear.
If the device runs on a battery
Portable AI devices, such as wearable recorders, desk companions and handheld assistants, add a power problem to the acoustic one. Listening for a wake word all day is a continuous load, so battery designs usually run wake-word detection on a low-power audio DSP or a codec with voice activity detection, and wake the main processor only when needed. Measure standby current with the microphones live, not just in sleep, because that figure sets the battery-life claim. A battery also brings lithium cell sourcing, a protection circuit, IEC 62133-2 safety testing and UN 38.3 for shipping, and at low charge the amplifier has less headroom, so loud playback can distort or cause a brown-out unless the power design allows for it.
Core manufacturing processes
- SMT, with care for MEMS microphones. They are sensitive to flux and liquids entering the port, and their makers generally advise against board washing or ultrasonic cleaning with the port exposed. A no-clean process and port protection are standard.
- Speaker assembly into the chamber with adhesive or gaskets, followed by an air leak test on sealed chambers.
- Fabric or grille assembly, where cosmetic quality and acoustic transparency both matter.
- Firmware and provisioning, including a unique device certificate for the cloud service.
- Acoustic end-of-line testing in acoustic test boxes: speaker frequency response, sensitivity, distortion and rub-and-buzz; sensitivity, frequency response and matching for every microphone; an echo test that plays a signal through the speaker and measures what the microphones pick up; and a wake-word check with recorded speech at a set level. These sit alongside the electrical functional tests.
Acoustic test boxes are expensive and slow compared with other stations, so line balance and the number of boxes need planning before the pilot production run.
Testing and validation from EVT to PVT
EVT: wake-word false-accept and false-reject rates in representative rooms with TV noise and reverberation; barge-in performance while music plays loudly; latency from the end of speech to the first audio response; thermal behavior of the processor and amplifier.
DVT: re-tune equalization and echo cancellation on tooled enclosures, because tooled plastic behaves differently from prototypes; run rattle sweeps, drop tests, temperature and humidity, ESD and long-duration speaker playback at high power; complete certification; and measure that the mute switch cuts the microphone signal in hardware, not only in firmware.
PVT: set acoustic test limits from the DVT distribution, check fixtures daily against a golden sample and a reference microphone, and track acoustic yield by station.
Certifications, privacy and AI rules
- Radio and EMC: FCC and ISED; CE under the Radio Equipment Directive, including its cybersecurity requirements for internet-connected radio products since August 2025; UKCA and the UK PSTI security rules; MIC, KC and RCM elsewhere.
- Safety and energy: safety testing to IEC 62368-1; efficiency rules for external power supplies; EU limits on standby and networked-standby power.
- Batteries: IEC 62133-2 and UN 38.3 for portable designs.
- Privacy: voice recordings are personal data under GDPR and many other laws; some US state laws, such as Illinois’ BIPA, treat voiceprints as biometric identifiers; products for children bring COPPA in the US and toy safety standards.
- AI rules: the EU AI Act requires that people be told when they are interacting with an AI system, unless it is obvious, and restricts emotion recognition in some settings. Confirm current obligations with a specialist.
- Ecosystems: qualification for any third-party voice assistant you integrate, and Matter certification if the device controls smart-home products.
Main cost drivers
- Processor and memory, especially for on-device models.
- Microphone count and tolerance grade.
- Speaker drivers, enclosure, and fabric or grille materials.
- Acoustic test equipment, cycle time and acoustic yield.
- Wake-word and audio front-end licensing.
- Power supply and certification.
- Cloud and model operating cost per user: outside the bill of materials, inside the business case.
What to ask a prospective OEM/ODM
- Which far-field voice products have you shipped?
- Who tunes the audio front end and echo cancellation, and across how many enclosure iterations?
- What acoustic tests run on every unit, with what equipment and limits?
- How do you match or calibrate microphones in production?
- Which microphone and speaker suppliers do you use, and what are their incoming limits?
- How do you protect microphone ports during SMT and assembly?
- How are device certificates provisioned, and who owns the cloud connection?
Common mistakes
- Proving voice performance on a development kit and assuming the enclosure will not change it.
- Tuning maximum volume for a showroom, which breaks echo cancellation at home.
- Buying wide-tolerance microphones for an array.
- Testing only “does it play and record” on the line.
- Implementing mute in software only.
- Pricing the device before modeling cloud cost per user.
Where Zefinity helps
We help teams decide the split between device and cloud, prepare the product for manufacturing and find OEM/ODM partners whose acoustic engineering and end-of-line test capability shows in products they have shipped. Through NPI, Launch support coordinates acoustic tuning rounds, DVT and pilot builds with the factory. See our approach to AI hardware.
Frequently asked questions
Should the AI run on the device or in the cloud?
Most products split it. Wake-word detection runs on the device; speech recognition and the language model often run in the cloud, which keeps hardware cost down but adds latency, an ongoing cost per user and a dependence on connectivity. Running more on the device needs a more capable processor and more memory, and changes the thermal design and the bill of materials.
How many microphones does a voice device need?
It depends on pickup distance and background noise. Close-talk devices and wearables can work with one or two microphones; far-field devices used across a room commonly use arrays of two to six or more, spaced to suit the beamforming algorithm. Extra microphones help only when they are matched, sealed and placed away from the speaker and its vibration.
Does a voice device need a hardware mute switch?
No single rule requires one in every market, but a switch that physically disconnects the microphones, with an indicator software cannot override, is common practice and the clearest answer to privacy questions from retailers and customers. Check the requirements of any voice platform you integrate, and test the switch on the production line like any other function.