PP-OCRv5 VoiceBridge — Snapdragon 8 Elite NPU (int8 TFLite)
PP-OCRv5 text detection and recognition for industrial ZH/KO safety signage, compiled to int8 TFLite via Qualcomm AI Hub for the Snapdragon 8 Elite (SM8750) Hexagon V79 HTP NPU.
Built for VoiceBridge — an offline VN↔EN/KO/ZH speech + sign translator for the OneVoice AI Challenge 2026 (Saigon AI Hub + Qualcomm).
Models
| File | Description | Size | NPU latency (S25 Ultra) |
|---|---|---|---|
pp_ocrv5_det_mobile.tflite | Text detection (PP-HGNetV2 + DB++) | 807 KB | 1.23 ms |
pp_ocrv5_zh_rec.tflite | Chinese text recognition (SVTR/CTC) | 22 MB | 0.99 ms |
pp_ocrv5_ko_rec.tflite | Korean text recognition (SVTR/CTC) | 3.9 MB | 0.49 ms |
pp_ocrv5_lat_rec.tflite | Latin text recognition (SVTR/CTC) | 2.2 MB | 0.31 ms |
zh_dict.txt | 18 383-char ZH dictionary | — | — |
ko_dict.txt | 11 945-char KO dictionary | — | — |
Full ZH/KO sign pipeline ≈ 2.2 ms on Hexagon NPU. 100% NPU — zero CPU fallback.
Performance
Latency — Samsung Galaxy S25 Ultra (SM8750, Android 15)
Measured via Qualcomm AI Hub cloud-hosted real device. Bench date: 2026-06-06.
Tool: qai-hub==0.50.0, qairt==2.45.0.
| model | NPU ms | NPU ops | vs ORT CPU |
|---|---|---|---|
| det | 1.23 | 156/156 (100%) | 105× faster |
| zh rec | 0.99 | 219/219 (100%) | 180× faster |
| ko rec | 0.49 | 223/223 (100%) | 86× faster |
| lat rec | 0.31 | 223/223 (100%) | 148× faster |
AI Hub job IDs: det=jg98xw1qp, zh=jp38we2x5, ko=jgnx7evm5, lat=jgj1wln8g
Accuracy — public dataset (ChineseOCRBench / ReCTS, Apache-2.0)
Real store-front signage photos (200 scored), standard ChineseOCRBench substring matching.
| metric | value |
|---|---|
| recall@full (answer fully recognized) | 48.5% |
| mean char-recall | 66.0% |
ReCTS is VQA-style (ground truth = one region; full-image OCR is substring-matched), so
recall@full understates pure recognition; char-recall reflects character-level quality on
real, cluttered, perspective-distorted signs. Eval script in the VoiceBridge repo
(benchmarks/ocr/bench_public.py).
Vietnamese: PP-OCRv5's Latin dictionary has no precomposed Vietnamese tone-mark vowels (verified by byte-level grep). Do not use this for Vietnamese — use a Vietnamese-specialized recognizer (e.g. VietOCR) or ML Kit.
Usage
# Download models
from huggingface_hub import hf_hub_download
det = hf_hub_download("nrl-ai/pp-ocrv5-voicebridge-snapdragon", "pp_ocrv5_det_mobile.tflite")
rec = hf_hub_download("nrl-ai/pp-ocrv5-voicebridge-snapdragon", "pp_ocrv5_zh_rec.tflite")
On-device (Flutter/Android + LiteRT + QNN delegate):
// See lib/pp_ocrv5.dart in vietanhdev/voicebridge for the full integration.
final engine = PpOcrV5Engine();
final results = await engine.recognize(imagePath, 'zh');
Source models & compilation
- Source: PaddleOCR 3.x (Apache-2.0) — https://github.com/PaddlePaddle/PaddleOCR
- ONNX exports: monkt/paddleocr-onnx (Apache-2.0, no pickle)
- Compiled:
qai-hub submit_compile_job(..., options="--target_runtime tflite --quantize_full_type int8")with random PTQ calibration data (100 samples per model)
Script: benchmarks/ocr/compile_aihub.py in https://github.com/vietanhdev/voicebridge
Citation
@software{voicebridge2026,
author = {Nguyen, Viet-Anh and {Neural Research Lab}},
title = {{VoiceBridge}: Offline On-Device Speech and Sign Translator},
year = {2026},
url = {https://github.com/vietanhdev/voicebridge},
note = {OneVoice AI Challenge 2026, Saigon AI Hub + Qualcomm}
}