Skip to content

v0.27.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 31 Aug 03:37
· 3 commits to main since this release
v0.27.0
57a896a

HotClip v0.27.0 — 先把人声听清，再谈包装

发布日期：2026-08-31

  • 48kHz 智能人声增强：降噪开关下新增显式“基础 / 智能 48k”档位。智能档首次按需下载约 10MB、按 SHA-256 校验的 DPDFNet2 高分辨率本地模型，不需要 Python，不上传音频。
  • 正确的成片音频顺序：智能增强只在多片段、跳剪与高潮前置全部组装完后执行，再交给音效与 BGM 混音；响度归一在增强之后计算，避免把包装层误当噪声。
  • 画质与同步保护：单声道/双声道分别增强，按带上下文的 30 秒块处理最长 180 秒成片；最终只重编 AAC 音频，视频流原样复制。模型尾帧不足 10ms 时用原声补齐，大于该范围则判失败，不冒险缩短时间轴。
  • 可用性优先：模型缺失、下载失败、解码异常或推理失败时自动执行原有“双高通 + 温和谱减”基础链；关闭和基础档保持既有行为。完成页、CLI、MCP 与 clips.json 如实区分智能生效、回退基础档和未应用。
  • 模型与诊断：模型注册表、设置页模型清单和桌面健康检查现在覆盖 11 类角色，智能人声模型与语音安全切点模型都能单独查看完整性和占用。
  • 本地多模态预设升级：新用户的可选 Ollama 视觉端点默认预填 qwen3.5:4b，一个模型同时处理文字与九宫格接触表；已有用户的模型选择不被覆盖。新增 pnpm quality:eval:vision，可在同一批样片、同一提示与解析路径上对比本地 VLM 的结构化成功率、耗时与画面证据。

验证：60 项聚焦回归通过，覆盖 PCM 分块/声道映射、短尾同步补齐、真实 FFmpeg 音轨替换、学习模型失败后的基础档回退、导出处理顺序、模型清单、CLI 与 MCP 协议。官方 DPDFNet2 48k 模型通过现有 sherpa-onnx 原生运行时实测，输出保持 48kHz，视频编码未变化。

HotClip v0.27.0 — Make the voice clear before adding the polish

Release date: 2026-08-31

  • Smart dialogue enhancement at 48kHz: the denoise control now exposes explicit Basic and Smart 48k tiers. Smart downloads a SHA-256-verified ~10MB DPDFNet2 high-resolution model on demand, with no Python and no audio upload.
  • Correct publish-audio order: learned enhancement runs only after jump cuts, stitched pieces and cold opens are fully assembled, then hands the result to SFX/BGM mixing. Loudness normalization is measured after enhancement so packaging is never mistaken for noise.
  • Picture and sync safety: mono/stereo channels are enhanced independently in contextual 30-second chunks for clips up to 180 seconds. Only AAC audio is re-encoded while video is stream-copied. A sub-10ms model tail is filled from original PCM; larger drift is rejected instead of shortening the publish timeline.
  • Availability-first fallback: missing models, failed downloads, decode errors or inference failure transparently run the exact historical double-highpass plus conservative spectral-subtraction chain. Off and Basic behavior remain unchanged. The completion view, CLI, MCP and clips.json report learned, fallback and skipped outcomes truthfully.
  • Model inventory and diagnostics: the settings inventory and desktop doctor now cover eleven model roles, including separate integrity/size visibility for speech-safe cuts and smart dialogue enhancement.
  • Updated local multimodal preset: fresh optional Ollama vision settings prefill qwen3.5:4b, sharing one model across text and nine-frame contact sheets without overwriting existing user choices. pnpm quality:eval:vision compares local VLM structured success, latency and visual evidence on the exact same fixtures, prompts and parser.

Verification: 60 focused regressions cover PCM chunking/channel mapping, short-tail sync preservation, real FFmpeg audio replacement, Basic fallback after learned failure, export ordering, model inventory, CLI and MCP protocol. The official DPDFNet2 model ran through the bundled sherpa-onnx runtime at 48kHz with the video codec preserved.

Assets 11
Loading