---

name: ocr-backends description: "Configure and troubleshoot VSE PaddleOCR models, language codes, thresholds, and CPU/CUDA/DirectML/ONNX backend acceleration." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

OCR Backends

Use this sub-skill for VSE OCR setup: PaddlePaddle/PaddleOCR installs, bundled PP-OCRv5 model selection, language codes, CPU/GPU/ONNX provider choices, recognition batches, confidence thresholds, and backend verification.

Read first

• model and language reference for language groups, Fast/Auto/Accurate model directories, and V5 names.
• backend compatibility for CPU, CUDA, DirectML, ONNX, and unsupported/new-GPU decisions.
• API reference for OcrRecogniser, SubtitleDetect, PaddleModelConfig, and HardwareAccelerator.
• troubleshooting for import/model/provider errors and low confidence.
• Run scripts/model_config_probe.py or scripts/hardware_backend_probe.py for safe read-only diagnostics.

Key decisions

• Choose CPU unless the user has a verified accelerator and backend-specific Paddle/ONNX packages.
• fast mode prefers mobile detection and mobile recognition for Chinese, Traditional Chinese, English, and Japanese; other languages use language group recognizers.
• auto and accurate use server detection; Chinese/English/Japanese use server recognition, while many non-CJK languages use grouped mobile recognizers.
• Do not treat a visible GPU as proof that VSE uses it. Verify Paddle compiled CUDA or ONNX providers.

Route elsewhere

• Extraction sequencing, VideoSubFinder, output/caches: extraction-workflows.
• GUI settings cards for these options: gui-batch-operations.
• Text cleanup and typo replacement after OCR: postprocessing-config.