Music Studio v0.1.4 · NVIDIA GPU / 显卡加速安装版

Music Studio v0.1.4 — NVIDIA GPU acceleration / 显卡加速安装版

Download Music-Studio-0.1.4-Setup-x64.exe, close Music Studio, and install over
the previous version. Songs and lyric corrections are stored separately and retained.

下载 Music-Studio-0.1.4-Setup-x64.exe，关闭 Music Studio 后覆盖安装。
歌曲和修改结果独立保存，无需卸载或删除旧歌曲。

Setup-x64.exe is an online installer. It automatically downloads about 2.20 GB,
supports resume and checks SHA256 before installation. Basic
analysis works offline after installation. For offline setup, keep the EXE and
both matching BIN assets in the same folder; assembly is automatic.

EXE 是在线安装器，自动下载约 2.20 GB 安装数据，支持续传和哈希核对。
离线安装时将 EXE 和两个同版本 BIN 文件放在一起，双击 EXE 即可自动安装。

• Python, PyTorch 2.5.1+cu121, CUDA 12.1 and cuDNN 9 are included.
• Demucs htdemucs separation, Whisper small lyrics and CREPE full melody use the
  NVIDIA GPU when Automatic passes the runtime check. No extra basic-model downloads.
• CUDA is checked with an actual tensor computation and Whisper compute-type support;
  an unavailable GPU is disabled with an explanatory message. CPU remains a slow fallback.
• Rhythm and harmony use librosa on CPU. Qwen and turbo remain optional source configurations.

默认“自动选择”会优先使用通过检查的 NVIDIA 显卡，人声分离、歌词识别和旋律提取均支持 GPU。
不需要另装 Python、CUDA Toolkit 或 cuDNN；需已安装兼容的 NVIDIA 显卡驱动。
节拍、和弦仍在 CPU 上运行。内置基础模型可离线使用。

Windows 10 22H2 / Windows 11 x64. RTX 4070 Laptop 8GB is the local validation target.
Other GPU families are not established by that check; AMD/Intel GPU acceleration is
not included. RTX 50-series support is not provided by this CUDA 12.1 runtime.
建议 16GB RAM，预留至少 10GB 安装空间。安装包尚未代码签名，请核对 SHA256SUMS.txt。

The small Windows.zip is source code, not a portable offline app.
体积较小的 Windows.zip 是源码包，不包含运行环境和模型。

Installation and actual inference evidence: GPU validation.
Composition and mixing belong to the independent Music Workbench.

Music Studio v0.1.3 · Standalone song analysis / 独立安装版

Music Studio v0.1.3 — 安装即用的歌曲分析 / Standalone song analysis

下载 Music-Studio-0.1.3-Setup-x64.exe，双击安装后从桌面打开 Music Studio。
无需 Python、终端、账号或 API。安装包内置 CPU 环境和基础模型，可离线分析自己的歌曲。

Download Music-Studio-0.1.3-Setup-x64.exe, install it, and open Music Studio.
Python, CPU inference dependencies and basic models are included. No terminal,
account, API key or extra basic-model download is needed.

• Demucs htdemucs：人声与伴奏分离 / vocal and backing separation.
• Whisper small：歌词和时间估计、改字 / timed lyrics and text correction.
• CREPE full：主旋律 MIDI / vocal melody MIDI.
• librosa：和弦与节拍信号分析 / chord and beat signal processing.
• 同步试听、跟唱、导出、删除与恢复 / synchronized playback, practice, exports and Trash.
• Independent bilingual high DPI desktop window and separate song storage.

CPU 基础版：Windows 10 22H2 / Windows 11 x64，建议 16GB RAM、4GB 安装空间。
整曲分析可能需要数分钟或更久，结果需核对。本版本没有新的准确率提升声明。
Qwen、Whisper turbo 和 GPU 为源码安装的可选配置，未包含在此安装器中。

The CPU edition is not real time. Review automatic results. Qwen, turbo and GPU
remain optional source installations. This release makes no new accuracy claim.

创作、编排、混音和 ACE-Step 属于独立的 Music Workbench。
早期工作台预发布已从此仓库公开下载中收起，历史标签保留。
Songs are saved in %LOCALAPPDATA%\MusicStudio\Songs and retained on uninstall.
Existing source workspaces are untouched; see Quick start.

核对 / Verify SHA256SUMS.txt. The app installer is unsigned. The Microsoft
WebView2 prerequisite is signature checked during packaging.

Demo: Let's Go Fishin' — Karissa Hobbs, CC BY 3.0, cropped and processed;
attribution.

Music Studio v0.1.2 · Preview

Music Studio v0.1.2 · Preview / 预览版

Fixes the Windows shortcut failing at lyric recognition when Qwen is selected.
修复双击启动后选择 Qwen，在歌词识别阶段报错的问题。

• The launcher automatically connects the separate environment installed by scripts/setup_lyrics.ps1.
• A custom environment can be selected with scripts/launch.ps1 -LyricsPython 'C:\path\to\python.exe'.
• Uninstalled recognizers are disabled in the bilingual UI. The API checks dependencies before doing work.
• Older running servers show restart instructions; existing songs, audio and result versions are preserved.

启动器自动连接已安装的歌词独立环境；未安装的识别模式在中英界面中停用。
后台提前检查依赖，避免分离音轨后才失败。正在运行的旧后台需重启，原有歌曲和结果均保留。

Download / 下载

Windows: Music-Studio-0.1.2-Windows.zip. Extract and double-click Start-Music-Studio.cmd.
Python: music_studio_local-0.1.2-py3-none-any.whl. Source: music_studio_local-0.1.2.tar.gz.
Verify downloads against SHA256SUMS.txt.

Requires Python 3.11 or 3.12. First startup installs preview dependencies.
The 45-second licensed demo works without neural weights. Analysis of your own songs needs:

.\scripts\setup.ps1 -Mode full -Device cuda -DownloadModels
# CPU: replace cuda with cpu / 无 NVIDIA 显卡时把 cuda 改为 cpu
# Optional Qwen / 可选 Qwen：
.\scripts\setup_lyrics.ps1
.\Start-Music-Studio.cmd

After updating or installing models, press Ctrl+C in the old startup console and relaunch.
Use the URL opened by the launcher; an older browser tab may still point to an old server.

升级或安装模型后，请在原启动窗口按 Ctrl+C，再双击启动。
使用启动器新打开的地址，旧标签页可能仍连接旧后台。

The ZIP does not bundle Python, CUDA or models. Lyrics, timing, melody and chords are estimates;
this patch improves runtime setup, with no new accuracy claim. See the
README and
validation.

ZIP 不含 Python、CUDA 和模型权重；本次修复启动与环境检查，不声称识别准确率提升。
示例：Let's Go Fishin' — Karissa Hobbs，CC BY 3.0，已裁剪、分离并压缩；
完整署名。

Music Studio v0.1.1 · Preview

Music Studio v0.1.1 · Preview / 预览版

First published preview. Includes the demo PCM-cache fix verified during release acceptance.
首个实际发布的预览版，包含发布验收发现的示例 PCM 缓存修复；0.1.0 构建已停止，未发布。

Unfold a song into vocals, backing, timed lyrics, vocal melody MIDI, chords and beats.
本地拆解歌曲，在同一条时间轴试听人声、伴奏、歌词、主旋律、和弦与节拍。

Download / 下载

• Windows: Music-Studio-0.1.1-Windows.zip. Extract and double-click Start-Music-Studio.cmd.
• Python package: music_studio_local-0.1.1-py3-none-any.whl.
• Source: music_studio_local-0.1.1.tar.gz.
• Verify file integrity against SHA256SUMS.txt.

Windows 用户解压 ZIP 后双击启动；Python 用户可安装 wheel；源码包用于开发。
使用 Get-FileHash -Algorithm SHA256 <file> 对照校验文件。

Requires Python 3.11 or 3.12. Initial startup downloads Python dependencies.
The ZIP is a source+wheel distribution, not a standalone EXE. No Python runtime,
CUDA toolkit, private recordings, benchmark audio or neural weights are bundled.

需预先安装 Python 3.11 或 3.12，首次启动联网安装依赖。ZIP 包含源码与 wheel，
不是独立 EXE；不包含 Python、CUDA、私人歌曲、测试集音频或神经模型权重。

Choose Try the demo / 加载示例歌曲 for a credited 45-second precomputed example:
immediate timeline exploration, original/vocal/backing playback, lyric correction,
sing along and export, without model downloads. Switch 中文 / English at top right.

To analyze your own songs, run:

.\scripts\setup.ps1 -Mode full -Device cuda -DownloadModels
# CPU: replace cuda with cpu / 无 NVIDIA 显卡时把 cuda 改为 cpu
.\scripts\launch.ps1

Included / 本版内容

Bilingual UI and documentation; stable lazy-loaded playback; editable lyric text;
review workflow and result versions; sing-along mode; recoverable project deletion;
43-case reproducible benchmark. See README.

中英界面与说明、稳定按需载入的试听、歌词改字与核对、结果版本与候选比较、跟唱视图、
歌曲删除恢复，以及有来源和哈希的 43 例评测。

Known limits / 已知限制

• Machine lyrics/timing/notes/chords need listening review. Gray text has line timing only.
• Main inference target: NVIDIA 8GB VRAM. CPU is slower; macOS/AMD/Intel GPU inference has not been validated.
• Qwen recognition requires a separate optional environment; dual timing guides are optional and can worsen timing tails.
• Long songs use more browser memory. There is no full notation or multi-instrument editor.

识别结果需试听核对，灰色文字仅有句级时间。主要推理验证环境为 NVIDIA 8GB；
CPU 较慢，macOS/AMD/Intel GPU 尚未验证。Qwen 需可选独立环境，双模型对齐可能扩大尾部误差。
长歌曲占用更多浏览器内存；暂不提供完整谱面或多乐器编辑。

Demo: Let's Go Fishin' — Karissa Hobbs, CC BY 3.0. Cropped, separated and
compressed; estimates are not ground truth. Attribution.