v0.26.0

HotClip v0.26.0 — 切得更紧，但一个尾音也不赌

发布日期：2026-08-31

• 本地语音活动证据：复用已经随应用提供的 sherpa-onnx 原生运行时，只新增一个 629KB、按发布方 SHA-256 校验的本地模型；不需要 Python，不上传音频。
• 安全自动边界：只有语音检测与逐字时间相互印证，才会在 0.6 秒硬上限内保护片头/片尾尾音，并把真实语音边缘交给后续镜头吸附作守卫。手动切点与拼接片外边界保持不动。
• 三重跳剪门：一段空隙必须同时无词、声学峰值低、语音检测为无人声才会删除。ASR 漏掉的轻声、短词或尾音会保留，笑声/掌声原有保护也继续生效。
• 多音轨一致：VAD、响度峰值、波形审阅与最终渲染显式使用探测选中的同一条音轨，不再让 FFmpeg 自动挑到另一条评论音轨。
• 可审计、可回退：clips.json 记录语音覆盖率、检测段数、边界位移和保护了几个剪口；区间过长、覆盖不足、模型/解码/缓存失败时完全沿用 v0.25 的切点逻辑。
• 模型完整性：模型注册表支持 raw 单文件 SHA-256 校验，损坏或被替换的模型不会被误判为已安装；桌面健康检查现在覆盖 10 类模型。

验证：新增纯函数、证据缓存、模型校验、跳剪保护与多音轨 FFmpeg 回归；真实带前后静音的系统合成语音经内置 FFmpeg 和现有原生运行时识别成功，临时模型与音频在测试后全部清理。

---

HotClip v0.26.0 — Tighter cuts, without gambling on a single phoneme

Release date: 2026-08-31

• Local speech-activity evidence: reuses the native sherpa-onnx runtime already bundled with HotClip and adds only one 629KB model verified against its published SHA-256; no Python and no audio upload.
• Safe automatic edges: only corroborated VAD/word evidence may protect outer speech within a hard 0.6-second movement cap, and the same speech edges constrain later shot snapping. Human-adjusted and stitched outer bounds never move.
• Three-gate jump cuts: a gap must be word-free, peak-quiet and VAD-negative before removal. Quiet words or phoneme tails missed by ASR stay, alongside the existing laughter/applause guard.
• Selected-audio parity: VAD, peak analysis, review waveforms and final render explicitly use the same probed audio stream instead of letting FFmpeg choose another track.
• Auditable exact fallback: clips.json records word coverage, span count, boundary deltas and protected gaps. Oversized windows, weak evidence, model/decode/cache failure preserve v0.25 behavior exactly.
• Model integrity: raw single-file model assets can now carry SHA-256 verification, so corrupt or substituted files are never treated as installed; desktop diagnostics now cover ten model roles.

Verification includes pure logic, evidence cache, model-integrity, jump-cut and multi-track FFmpeg regressions plus a real synthesized-speech smoke through bundled FFmpeg and the existing native runtime. All temporary model/audio files were removed after the smoke.