中文
本版更新分离引擎、默认伴奏、多轨试听和跨平台运行包，并首次提供 CPU / CUDA Docker 镜像。
主要更新
- 修复旧 UVR5 频谱缓冲区未初始化导致的偶发 NaN 音频；内置实现和独立固定官方运行时均初始化未使用频点，保留模型、复数精度和重采样参数。新增确定性回归，完整套件为 343 项。
- 补齐新版 PyTorch 间接引入的 NVIDIA 动态库，修复 Linux CUDA 便携包在源约束阶段缺少 NVRTC 内置库的错误。Windows/Linux GPU 构建均增加真实 NVRTC 编译检查，1.5.0 草稿与早期 1.5.1 候选包未公开，所有正式文件从本版标签重新生成。
- 分离引擎升级到 audio-separator 0.47.0。默认纯伴奏采用 Leap Instrumental 62，人声使用 Leap XE 90；保留 MVSep 9205 三模型主唱/和声分离，去混响采用 Stereo De-Reverb 22.5050。移除 PolarFormer 入口和发行权重。
- 翻唱结果提供同步多轨时间轴，支持静音、独奏、逐轨音量、时间偏移、缩放和拖入对照音轨；七类原始 WAV 可分别试听、下载。前端资源随包提供，可离线使用。
- 修复新版官方 RVC 目录结构变化导致的启动/转换失败（#4）。VC 与 UVR5 使用分别固定并逐文件校验的上游源码，完整包内置所需源码与基础模型。
- 校验角色模型的真实架构、说话人、音高条件与索引维度；修复 FAISS 稀疏近邻、精确重复向量和旧缓存复用问题。下载或导入失败保留原有资产并报告原因。
- 同步 Windows、Linux、WSL、Colab、Space 的依赖和文档，修复 Windows 依赖文件编码、冻结工作进程、Apple Silicon TorchAudio 版本组合及上传/下载问题。
- Docker 提供非 root 运行、持久化数据卷、可选密码文件、健康检查和反向代理子路径；设备不可用时明确停止。模型、精度、重叠参数和 Float32 音频输出策略与本地应用一致。
下载与升级
| 使用环境 | 选择 |
|---|---|
| Windows + NVIDIA 显卡 | AI-RVC-Windows-GPU-Portable.7z.*，内置 CUDA 12.8 运行库 |
| Windows CPU | AI-RVC-Windows-CPU-Portable.7z.* |
| Linux x86-64 + NVIDIA 显卡 | AI-RVC-Linux-GPU-Portable.tar.gz.part* |
| Linux x86-64 CPU | AI-RVC-Linux-CPU-Portable.tar.gz.part* |
| Docker | ghcr.io/mason369/ai-rvc:1.5.1-cuda 或 ghcr.io/mason369/ai-rvc:1.5.1-cpu |
下载所选包的全部同前缀分卷和对应 SHA256SUMS。Windows 用 7-Zip 打开 .7z.001 解压；Linux 按文件名顺序合并分卷后解压。请解压到新目录，再迁移自己的角色模型与输出文件，避免旧运行库混入。便携包内置基础模型、默认六个分离模型、官方源码、FFmpeg 和播放器；角色模型按需下载或导入。
Docker 使用仓库提供的 Compose 和单一数据卷，升级时保留数据卷。详细命令及首次启动、密码、备份、代理配置见 Docker 使用指南。CPU 与 CUDA 镜像仅声明 linux/amd64。
验证范围
四种正式便携包均使用同一段五秒真实音乐完成默认与官方路线，共验证 44 个 Float32 输出；四个平台的 Actions 构建各通过 343 项测试。GPU 实测为 RTX 4070 Ti SUPER 16 GB、驱动 595.97；Linux 与 Docker 使用 WSL2 Ubuntu 22.04，Windows 宿主可见物理内存为 31.10 GiB。其他 NVIDIA 显卡并未因此计为通过。
两种 GHCR 正式镜像均已公开，发布作业各通过 343 项测试；随后使用空白登录配置匿名拉取，核对摘要和版本，并以 UID 1000 断网完成默认与官方路线，另验证 22 个 Float32 输出。首次 CUDA 拉取曾遇 WSL DNS 超时；连通性恢复后再次拉取通过，没有修改系统 DNS、代理或推理参数。完整记录见验收附件。
容量规划请参考 README 的默认完整链建议：NVIDIA 16 GB 显存、64 GB 系统内存。它是整曲运行的规划参考，不是本次验证出的最低配置。四种便携包实际解压大小为 4.53–12.44 GB，还需同时容纳下载分卷、自己的角色模型、输入与输出；逐包压缩及解压字节数见验收附件。
正式发行资产的逐项 SHA-256、文件清单、实际运行和旧版对照结果见随附 RELEASE-VALIDATION.md。Windows、Ubuntu、Apple Silicon 的依赖/合同 CI 与真实 GPU 翻唱是不同层次的证据；DirectML 不支持默认 RoFormer 链，未接入的 ROCm/XPU/MPS 设备、云端完整翻唱和手机真机仍不计为验收完成。
模型对比属于受控样本测试，不能证明所有歌曲均提升。升级不降低模型、推理精度或输出质量参数。完整改动见 CHANGELOG。发行标签和程序版本保持 v1.5.1；文档提交 efb047e 补充发布后的实际验收结果。
关于新报告的 CVE-2026-28414：官方触发条件是 Windows + Python 3.13 及以上；本版固定 Python 3.10，Docker 运行 Linux。当前 Windows 3.10 环境的路径与自建文件 HTTP 检查均拒绝越界访问。Gradio 仍为 5.49.1，不能将此解释为已升级依赖或所有漏洞都已排除；适用范围与后续跟进见 issue #5及平台验收文档。
English
This release updates separation, default instrumental stems, multitrack playback and portable runtimes, and adds CPU / CUDA Docker images.
- Fix intermittent NaN audio caused by uninitialized UVR5 spectrum bins in both bundled and pinned upstream implementations. Preserve models, complex precision and resampling parameters; deterministic regression is included in the 343-test suite.
- Include indirectly required NVIDIA runtime libraries and validate actual NVRTC compilation in GPU builds. This fixes missing NVRTC builtins during source-guided reconstruction in Linux CUDA portable packages. The v1.5.0 draft and early v1.5.1 candidate packages were rejected; final artifacts are rebuilt from this release tag.
- Upgrade to audio-separator 0.47.0. Use Leap Instrumental 62 for pure accompaniment, Leap XE 90 for vocals, the MVSep 9205 three-model lead/backing ensemble and Stereo De-Reverb 22.5050. Remove PolarFormer runtime entries and bundled weights.
- Add synchronized multitrack playback with mute, solo, gain, offsets, zoom and reference tracks. Keep all seven original WAV outputs available separately and bundle frontend assets for offline use.
- Fix the changed upstream RVC layout (#4). Pin and validate separate VC and UVR5 source trees, and include them and their base models in portable packages.
- Strengthen model/index validation, preserve native model contracts and fix sparse/duplicate FAISS retrieval and stale validation caches. Failed imports or downloads preserve existing assets and report the cause.
- Align dependencies and documentation across local and cloud entry points, including Windows encoding, frozen workers, Apple Silicon TorchAudio compatibility and file transfer fixes.
- Provide non-root Docker images with a persistent data volume, optional password-file authentication, health checks and reverse-proxy support. Keep the same models, precision, overlap settings and Float32 output policy; unavailable devices fail explicitly.
Download every volume for your chosen platform/variant and its SHA256SUMS. Open .7z.001 with 7-Zip on Windows, or concatenate the Linux .part* files in filename order before extraction. Extract into a fresh directory and migrate your own character models and outputs. Base models, all six default separator weights, pinned upstream sources, FFmpeg and the player are included; character models are downloaded/imported separately.
Docker images are ghcr.io/mason369/ai-rvc:1.5.1-cuda and ghcr.io/mason369/ai-rvc:1.5.1-cpu for linux/amd64; use the supplied Compose files and retain the data volume when upgrading. See the Docker guide.
All four distributed portable apps completed both routes with the same five-second music sample, producing 44 valid Float32 outputs; each build passed 343 tests. GPU acceptance used an RTX 4070 Ti SUPER 16 GB, driver 595.97; Linux/Docker ran on WSL2 Ubuntu 22.04, with 31.10 GiB of visible physical RAM on the Windows host. README planning guidance for full songs is 16 GB of NVIDIA VRAM and 64 GB of system RAM, not a verified minimum. Portable packages unpack to 4.53–12.44 GB; allow additional room for downloaded volumes, character models and audio.
Both GHCR images are public. Each publication job passed 343 tests, followed by anonymous pulls with empty credentials and network-disabled default/official inference as UID 1000, validating 22 additional Float32 outputs. The initial CUDA pull failed on WSL DNS; a subsequent pull passed after connectivity checks without changes to DNS, proxy or inference settings.
See the attached RELEASE-VALIDATION.md for exact distributed-artifact checks, real runtime evidence and comparison with v1.4.1. CI installation/contract checks are distinct from real GPU inference. DirectML does not support the default RoFormer chain; unavailable ROCm/XPU/MPS hardware, complete cloud covers and physical mobile devices remain unverified. Controlled model comparisons do not establish improvement for every song. See the full changelog. The release tag and application remain v1.5.1; documentation commit efb047e records post-publication acceptance.
The newly reported CVE-2026-28414 requires Windows and Python 3.13+. This release uses Python 3.10, and Docker runs Linux. Local Windows 3.10 path checks and an owned-file HTTP probe rejected traversal. Gradio remains 5.49.1; this is an applicability finding, not a dependency upgrade or proof that all dependencies are vulnerability-free. Follow issue #5 for the remaining report and compatibility work.