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.
From capture to "should this be uploaded?" — the edge does it in one pass; the cloud only receives speech worth recognizing.
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.
→Encrypted WebSocket, cloud-grade LLM recognition.
EBOX uses a new-generation open-source LLM-architecture recognition model.
| Model | Languages | Strengths & limits | Latency |
|---|---|---|---|
| Qwen3-ASR-1.7BUsed by EBOX | 30 languages + 22 Chinese dialects zh/en/ja/ko/vi/th/de/es/pt/id全覆盖 | LLM-architecture recognition; 10k-token industry context biasing — the rarer the SKU or batch number, the better it works; real-time inference on pure CPU, no GPU needed | Streaming recognition while speaking 0.1–0.3s per second of audio full text when you finish |
| Legacy local ASR Paraformer class | Mostly zh/en | Mature architecture, but limited language coverage and weak hot-word expansion | Fast, but languages limited |
| Whisper | 90+ languages | Strong general-purpose; no hot-word mechanism, hallucination risk, heavy compute | Hard to run real-time on CPU |
| Cloud API | Broad coverage | Depends on external network, voice data leaves the plant, pay-per-use forever | Affected by network jitter |
Local synthesis engine — clean licensing, no cloud dependency.
| Model | Languages | Strengths & limits | Latency |
|---|---|---|---|
| Piper / VITS local synthesisUsed by EBOX | 60+ languages Vietnamese with 3 voices at launch | Fully local CPU synthesis, MIT-family license safe for commercial use; one runtime, engines added per language; front number/bin-code reading-normalization layer for unambiguous announcements | First packet 100–300ms streaming playback, speaks while generating |
| MeloTTS | 6 languages | CPU real-time, zh/en mixed reading — a multi-language backup engine | Real-time |
| Cloud TTS | 140+ languages | Great quality but needs internet — unavailable on industrial intranets | Affected by network RTT |
Book a demo — bring your own industry hot words to try.
