Selecting Edge Voice AI Hardware: Transcription, Conversation, Pickup and Compute

Voice AI hardware follows from three decisions: transcription or conversation, what must run locally, and where the microphone goes. They determine, in turn, whether a speaker and on-board echo cancellation are needed, which compute tier the box sits in, and which microphone model and connection mode to use.

This guide sets out a four-step selection method, with the form factors, rated pickup ranges and reference prices of reSpeaker AI microphones, and measured transcription concurrency, conversation languages and local LLM options for 5 local compute devices. Channel counts are measured on the ASR service with a fixed corpus and exclude VAD, upload and storage — see the measurement basis at the end.

Quick answer

The table below summarises the conclusions derived later. It maps one-to-one to the three questions in the configurator on the Acoustic & Audio AI Tech page (what to do / what runs locally / where the microphone goes).

Reference prices are for the compute device at the time of lookup and exclude the microphone (microphone prices are in Step 2); the product page is authoritative when ordering.

Three conditions change the table above:

• Barge-in required — the microphone is limited to reSpeaker XVF3800 or reSpeaker Flex, whose on-board AEC removes the device's own playback; the configurator's barge-in routes do not include Lite
• Conversation languages beyond Chinese and English — the reComputer RK3576 conversation pipeline covers Chinese and English; for other languages choose reComputer RK3588, J3011 or J4012, which cover all 30 conversation languages
• Speaker more than ~3 m from the microphone, or steady noise above ~70 dB — adjust microphone position or count before sizing compute; recognition accuracy degrades under these two conditions before compute becomes the limit

If you already know the job, the local scope and the microphone position, use the configurator below to generate the device list; to see the basis for each choice first, read the four steps.

---

Why voice processing runs locally

"What runs locally" is an explicit option in the four steps below. The choice rests on three properties of local processing.

Audio stays on site. With local transcription, VAD and ASR run on site and only the final text of each utterance is uploaded. With speech local and the LLM in the cloud, only the recognised text leaves the site.

Offline operation after setup. On the fully local conversation route, once the first online start has downloaded the images and models, VAD, ASR, TTS and the LLM all run on the device without an internet connection.

Same interface as the cloud. Both local LLM routes (reComputer J4012, and reComputer RK3588 with an NPU expansion card) expose an OpenAI-compatible Chat Completions endpoint on the device. Switching between local and cloud changes only the endpoint address; the application above it is unchanged.

---

System anatomy and three voice capabilities

AI microphone → compute box → speaker / output → application

A voice system has four roles:

• AI microphone — picks up sound and runs front-end processing on the board (acoustic echo cancellation (AEC), noise suppression, beamforming) before handing audio to the host
• Compute box — runs whatever must stay local: voice activity detection (VAD), speech recognition (ASR), speech synthesis (TTS) and the large language model (LLM)
• Speaker / output — conversation routes play replies through a speaker connected to the microphone board, so playback and echo cancellation share one board; transcription routes output text only
• Application — receives text or voice and connects to business systems

By interaction type, voice applications fall into three classes, each needing more capability than the last.

Transcription: turn speech into searchable text

Conversations at store counters, meeting rooms, clinics and inspection sites are converted into structured text for review and search. The device does not reply, so TTS, an LLM and a speaker are not needed. What is needed is VAD (cutting out segments where someone speaks) and ASR.

Conversation: understand and answer by voice

At hands-busy workstations, voice queries and replies replace screen interaction

The device hears a question, an LLM generates the reply, and TTS plays it back, one turn at a time. On top of the VAD and ASR used for transcription, this adds TTS, an LLM and a speaker.

Barge-in conversation: users can interrupt while the device speaks

While the device is playing a reply, the user starts speaking and the device stops playback at once to listen to the new utterance. The microphone must still separate the human voice while the speaker is playing, which depends on AEC on the microphone board. Do not mute the microphone during playback to avoid echo: the echo disappears, and so does the ability to interrupt.

---

Step 1: Fix the job and derive the capabilities it needs

Two rules:

1. Text only, the device never replies: take the transcription route. Step 2 allows the wearable Clip, and Step 3 allows no compute box (cloud transcription)
2. The device replies: the microphone board needs speaker output and on-board AEC; if users will talk over the device while it speaks, Step 2 is limited to XVF3800 or Flex

Step 2: Fix the pickup and connection mode

Choose the model by where the microphone goes

Four selection rules:

1. A room with several people, or barge-in needed: choose XVF3800. 360° pickup, on-board AEC and noise suppression, and one USB-C cable for power and data; it is the default microphone of the conversational voice reference design
2. The microphone goes inside a robot or product enclosure: choose Flex. The array board and processing board are separate and joined by a ribbon cable, so the array can sit anywhere in the enclosure; for front-facing pickup only, choose Linear-4. 10 W speaker loads are driven through the board's external 12 V power terminal
3. One or two people at a desk, transcription or conversation without barge-in: choose Lite
4. The speaker moves around (inspection, field work, in-store assistance): choose Clip. It records for 14–18 hours without a phone nearby and uploads audio through the phone app; transcription only

Choose the connection by where the host sits

• Microphone and compute box in one place: connect over USB and buy the version without XIAO, which works out of the box
• Microphone far from the host, or only an MCU in the product: buy the version with XIAO ESP32S3 and flash the I2S firmware; the server-side audio receiver and transcription must be deployed separately, so delivery effort is higher than on the USB route. Firmware and example projects are on the reSpeaker XVF3800 Wiki
• Clip uses neither mode; audio is uploaded through the phone app

Scope. Pickup range and coverage are rated values; actual range on site depends on reverberation and noise, checked in Step 4.

Step 3: Decide what runs locally and choose the compute tier

Local scope options

Measured results by compute tier

Test conditions: SenseVoice model, each clip no longer than 4 seconds, p50 / p95 timed from ASR end-of-speech to final result, excluding VAD, upload and storage. Accuracy was evaluated separately on one 100-item corpus: Chinese CER is 4.82% on all 4 devices, and English WER ranges from 7.50% to 8.51%. Reference prices are whole-unit prices at the time of lookup; the product page is authoritative.

Four selection rules:

1. Transcription only, single site, one or two speakers: reRouter CM4 is sufficient; for multiple sites or concurrent channels choose reComputer RK3576
2. Conversation with a cloud LLM: reComputer RK3576 for Chinese and English; reComputer RK3588 for other languages
3. Fully local conversation: choose reComputer J4012; for projects already on the RK platform, reComputer RK3588 with an NPU expansion card (contact a solution expert)
4. The same device also runs vision or other AI workloads: choose reComputer J3011, with the LLM in the cloud

Scope. The table is a side-by-side comparison on one corpus and one ASR service, for narrowing the selection range; it is not an acceptance metric. Test on the target site when: individual utterances are well over 4 seconds; VAD, upload and storage must be counted in end-to-end latency; the main language is English or a language other than Chinese and English; or reRouter CM4 is used for formal capacity planning.

Step 4: Check site acoustics

Compute sizing only holds if the microphone receives recognisable speech. Check in this order: distance and noise first, then echo.

Beyond ~3 m or above ~70 dB of steady noise, the array can no longer separate the speaker, and word error rate degrades before compute has any effect. These two boundaries come from the scope notes of the retail voice transcription reference design; they serve as site-survey guidelines rather than test thresholds.

Acceptance test for barge-in conversation: set the speaker to normal room volume, run three consecutive rounds in a real room, interrupt 0.5–1 second after each reply starts, and confirm that the old reply stops at once and the interruption is not lost. A terminal that passes at low volume can fail at working volume for purely acoustic reasons, unrelated to the model.

---

For solution providers and system integrators: from proof to delivery

If the solution is delivered to downstream customers, confirm the following three items at the selection stage.

Proof path. Conversation and barge-in conversation have a runnable reference design for building a demo during the proposal stage: see Conversational voice AI.

Integration with your own systems. The integration point for conversation is an OpenAI-compatible streaming Chat Completions endpoint: point the endpoint address at your own RAG service, ticketing or ordering back end, or robot command layer, and the voice layer needs no changes. Switching between local and cloud LLMs likewise changes only the endpoint address. The transcription route uploads the final text of each utterance to the business system.

Hardware customisation and volume supply. Enclosure, branding, packaging and interface changes can be evaluated as a customisation project; the split structure of reSpeaker Flex lets the array be placed to suit the product enclosure. For customisation scope, certification coverage and lead times, see ODM/OEM customization services.