ASD Pipeline
把 Active Speaker Detection 变成可恢复、可测试、可复用的本地能力入口A recoverable, testable, reusable local entry point for Active Speaker Detection
Overview
这是什么What It Does
一个面向 Apple Silicon 本地验证的 Active Speaker Detection 工作台，把 LR-ASD 从一次性 CLI 运行整理成分阶段、可恢复、可输出结构化工件的基础能力。An Apple Silicon workspace that turns LR-ASD from a one-off CLI run into a staged, resumable pipeline with structured artifacts.
面向 视频理解实验、说话人分析、字幕与多模态工作流开发者。Built for developers working on video understanding, speaker analysis, subtitles, and multimodal workflows.
更多说明More Details
原始的 ASD 基线常常更像一次性研究脚本：整跑重、恢复弱、中间状态不透明。这个项目的价值不在于换了一个新模型，而在于把整条流程拆成清晰阶段，并把结果落成结构化 JSON。
每次运行会输出 tracks.json、frame_scores.json、predictions.json、metrics.json 和 validation_overlay.mp4，比单纯”跑出一个可视化视频”更适合作为下游能力模块。
Many ASD baselines behave like one-off research scripts: complete runs are expensive, recovery is weak, and intermediate state is opaque. This project makes every stage explicit and stores results as structured JSON.
Each run produces tracks, frame scores, predictions, metrics, and a validation overlay, making the pipeline useful as a reusable downstream capability.
Features
核心能力Core Capabilities
把 extract、detect、track、score、render 拆成显式阶段，便于局部调试和定位问题。Extract, detect, track, score, and render are explicit stages that can be inspected and debugged independently.
支持从中间工件恢复、跳过已有阶段、强制重跑指定阶段，适合离线实验流程。Resume from intermediate artifacts, skip completed stages, or force selected stages to run again.
统一产出 tracks、frame scores、predictions 等 JSON 工件，便于后续分析和集成。Tracks, frame scores, predictions, and other JSON artifacts are ready for analysis and integration.
