Treating OCR as “one model” is a common misconception. Results depend on the whole pipeline: detect, rectify, recognize, structure. OCR is a pipeline, not a model Text detection: DBNet-class models find text lines; skew-correct perspective first; Recognition: quantized CRNN/SVTR on the NPU runs 20-50 ms per line; Structuring: template-based field extraction (invoice no., date, amount) […]
Vendor-proprietary protocols make integration expensive. OPC UA offers a common information model and a secure channel, the lingua franca of cross-system integration. Why OPC UA Information modeling: don’t expose bare variables — build a device/unit/station object tree so MES understands data semantics; Security: use certificate mode, not anonymous, and rotate X.509 certificates on a schedule […]
Industrial voice doesn’t need to sound human; it needs to work offline, with stable latency, forever. Lightweight TTS like Piper fits exactly. What industrial TTS actually needs Latency first: Piper synthesizes a sentence in 100-300 ms on CPU or NPU — fine for alarms; Bake voices offline: ship fixed voice models, avoiding runtime downloads and […]
Home Assistant’s voice stack splits into wake word, speech-to-text and intent handling. RK3588 has the headroom to keep all three local. Three local stages microWakeWord for wake: enabled directly as an add-on with low CPU cost; Whisper for STT: the small model transcribes a sentence in 1-2 s on RK3588; Chinese accuracy is acceptable; Piper […]
RK3576 carries the same 6-TOPS NPU as RK3588; the differences are CPU and peripherals. Single-stream vision is fully within reach. Single stream is the comfort zone Measured levels: YOLOv5s @640 about 25-30 fps, face detect+recognize 15 fps, plate OCR 20 fps — all smooth for one stream; CPU is the bottleneck: multi-stream decode loads the […]
Whisper is entirely usable on RK3588, but offline batch and streaming have very different expectations. Batch works; realtime takes conditions Offline batch: the small model transcribes one hour of audio in 15-25 minutes; faster-whisper INT8 goes faster still; Streaming: needs VAD segmentation plus chunked inference with 3-8 s latency — fine for subtitles, not interpretation; […]
The customer didn’t want to replace the line, only to add automatic inspection at key stations and reduce escape rates from manual checks. Retrofit, don’t rebuild Deployment: one edge box per station with an industrial camera, results driving a reject mechanism directly over IO; Algorithm: trained on historical defect images, with edge inference plus human […]
RK3588’s NPU has three cores that run independently or together. The core question in multi-task scenarios: which model gets which core. The art of core assignment Assign by priority: the main detector owns core0; secondary classifiers/OCR share core1/2; Avoid frequent switching: multi-model switching on one core costs context overhead; stagger by business peaks; Combined mode: […]
Not every job belongs on the main SoC. Hard-realtime control on an MCU with perception and scheduling on the host is the sturdier design. Task split matters more than TOPS Layered architecture: motor loops and safety stop on an MCU; SLAM and visual avoidance on the RK3588 host; Compute allocation: lidar SLAM is CPU-bound, visual […]
Quantization compresses FP32 weights to INT8 with big speed and memory wins, but a calibration set mismatched to your scene collapses accuracy. The calibration set is the soul of quantization Calibrate from real cameras: 200-500 frames covering day/night, weather and angles beat generic datasets; Per-layer sensitivity: keep accuracy-critical head/tail layers in FP16 mixed quantization; Always […]