Devices that hear on their own, conversations as natural as speech, and solutions that land on the floor — the technology base behind EBOX × EAR CLOUD.
An employee asks aloud; the EBOX terminal senses in milliseconds and answers in full duplex; EAR CLOUD handles recognition, understanding and enterprise-system calls. No walking to a PC, no hands, no training.
system APIs wrapped by the MCP layer — not one line of your code changes
wake / recognize / understand / synthesize all on the intranet — voice data never leaves the plant
rules → LLM → on-device commands; the line never goes deaf
An employee speaks, an answer comes back in seconds — powered by three technologies working together: on-device models make the terminal hear fast, full-duplex makes the conversation natural, integrated engineering makes it land.
Wake, denoise and recognition models run on the terminal chip — hearing without depending on the network.
Full breakdown →Speak while it listens, barge in anytime — as natural as talking to a colleague.
Full breakdown →Fully in-house from board design to cloud platform, delivered as one integrated product.
Full breakdown →The first gate on an industrial site sits on the terminal chip: wake, denoise and detection all run locally — only valid speech goes upstream. Noise-resistant, bandwidth-saving, privacy-preserving.
Custom brand wake word with a TTS-synthesis training pipeline; accuracy 95-98% of real human recordings.
AEC echo cancellation + noise suppression + array beamforming, locking onto the voice amid noise.
Neural VAD separates speech from pauses at millisecond level.
30 on-device commands work without network — the line never goes deaf.
In productionAudio streams upstream over encrypted WebSocket to EAR CLOUD: Qwen3-ASR-1.7B recognition (30 languages + 22 Chinese dialects, 10k-token industry context biasing, real-time on pure CPU), plus local Piper synthesis (60+ languages, first packet 100-300ms). Per-frame on-device processing takes 2-3ms with zero false-wake audio upstream.
Traditional voice terminals are half-duplex — you cannot cut in while the device speaks. EBOX achieves true interrupting full duplex with three layers: cut in anytime during playback and the device goes quiet within 100ms to listen.
Neural VAD distinguishes speech / pause / noise in 75dB plant noise, sensing who speaks at millisecond level — the first reaction to an interruption comes from the edge, not the cloud.
Linear + non-linear dual-filter AEC cancels the device's own playback echo with 40dB suppression — it knows what it is saying and never mistakes its own voice for a command.
A 160ms acoustic window continuously judges dialogue state (listening / speaking / pausing / idle), with turn decisions in about 250ms — telling thinking from finished, at human pace.
When an employee cuts in mid-playback: muted within 100ms, relistening, re-answering in seconds — interruption, correction and takeover all feel natural.
SoulX-Duplug is open-sourced (Apache-2.0) by Soul AI × Shanghai Jiao Tong University × Northwestern Polytechnical University, built on voice technology proven across hundreds of millions of users.
EBOX-200 is built on Espressif's new-generation flagship ESP32-S31; the EAR CLOUD platform runs distilled lightweight models — edge-cloud synergy with no GPU server required.
Espressif's new dual-core RISC-V chip entering mass production in 2026: 320MHz, 128-bit SIMD per core, AI instruction acceleration — about 65% more CoreMark performance than the previous S3.
250MHz DDR PSRAM, up to 64MB SiP; dual I2S controllers for hardware-level audio sync, low latency and high fidelity.
802.15.4 and Gigabit Ethernet MAC for industrial RF environments; 60 GPIOs, DVP camera and LCD reserved for product evolution.
Secure boot + Flash/PSRAM encryption + AES/RSA/ECDSA hardware acceleration.
Platform side: distilled Qwen3-4B understanding model and Qwen3-ASR-1.7B recognition model, quantized and running real-time on pure CPU (one 16-core/32GB server carries 5 concurrent sessions); dual-route NLU — frequent questions answered in about 1 second via a rule engine straight to enterprise systems, long-tail questions handled by the LLM; an MCP integration layer standardizes system calls so not one line of your code changes.
These three are not parallel buzzwords — they form one complete interaction chain:
Voice is woken up and captured locally — response starts as you speak.
Speak while it listens, barge in anytime — one sentence gets it done.
In-house terminal + cloud platform — deploy on site, connect systems, get to work.
One software stack, four deployment shapes. Choose by scale and budget — start light, grow smoothly; upgrades add hardware only, never change the architecture or require a reinstall.
| Plan | Composition | Hardware | Scale |
|---|---|---|---|
| LightweightPilot pick | Voice algorithms, platform and LLMs co-deployed on one regular server — a single machine carries the whole chain | 1 × regular server 16-core / 32GB / 500GB · no GPU | ≤10 EBOX units 5 concurrent peak |
| Industrial | Voice algorithms pushed down to edge boxes (shop-floor side); platform & LLMs on a regular server — algorithms close to the field | N × edge boxes (1 per workshop) + 1 × regular server | Unstable workshop networks offline fallback matters |
| StandardMain pick | Algorithms & platform on one server, inference LLMs on another — separated compute, isolated fault domains | 2 × regular servers LLMs own the whole machine | 10–30 EBOX units 5–15 concurrent |
| High-performance | Platform on a regular server; LLMs & algorithms on a GPU server (vLLM batch inference acceleration) | 1 × regular server + 1 × GPU server | 30+ units dozens of sessions · multi-site |
All four tiers run the same stack: upgrades only add servers / edge hardware — switch config to migrate, zero business interruption.
First scenario: manufacturing warehousing — inventory, bin locations, inbound/outbound, one sentence away.
