SXSEditor

Quick Start 快速开始

This guide walks you through the complete workflow: from installing SXSEditor to synthesizing your first singing voice. Follow each step in order.

本指南将引导你完成完整工作流程：从安装 SXSEditor 到合成第一段歌声。请按顺序逐步操作。

ℹ️

SXSEditor is a desktop application for singing voice synthesis (SVS). It uses the SoulX-Singer neural model on ONNX Runtime with DirectML GPU, WebNN NPU, and CPU support. Supported singing languages: English, Chinese (Mandarin), and Japanese (via English phoneme migration or hybrid phoneme mapping on the base model; JP LoRA model in development).

SXSEditor 是一个歌声合成（SVS）桌面应用。基于 SoulX-Singer 神经网络模型，通过 ONNX Runtime 运行，支持 DirectML GPU、WebNN NPU 和 CPU 推理。支持的合成语言：英语、中文（普通话） 和 日语（通过英语音素迁移或混合音素映射在基础模型上合成；日语 LoRA 模型开发中）。

System Requirements 系统要求

Component组件 Minimum最低 Recommended推荐
OS操作系统 Windows 10/11 64-bitWindows 10/11 64 位 Windows 11 64-bitWindows 11 64 位
CPUCPU Any modern multi-core任意现代多核处理器 Recent Intel/AMD 6+ cores较新的 Intel/AMD 6 核以上
RAM内存 8 GB8 GB 16 GB16 GB
GPUGPU None (CPU fallback)无（CPU 回退） Discrete GPU with 4GB+ VRAM独立显卡，4GB 以上显存
Storage存储 3 GB (app + models)3 GB（应用 + 模型） SSD with 10+ GB freeSSD，剩余 10 GB 以上

A discrete GPU is strongly recommended for reasonable synthesis speed. CPU inference works but is significantly slower.

强烈推荐使用独立显卡以获得合理的合成速度。CPU 推理可用但速度明显较慢。

Step 1: Download & Install 第 1 步：下载与安装

Download the Windows installer (models are auto-downloaded on first launch):

下载 Windows 安装包（模型在首次启动时自动下载）：

  1. Run the installer. Administrator privileges are required for installation.运行安装程序。安装需要管理员权限。
  2. Launch SXSEditor from the Start Menu or desktop shortcut.从开始菜单或桌面快捷方式启动 SXSEditor。
💡

macOS / Linux users: build from source. See the Developer Guide for instructions.

macOS / Linux 用户：请从源码构建。详见开发者指南。

Step 2: Download Models (First Launch) 第 2 步：下载模型（首次启动）

On first launch, SXSEditor detects that inference model files are missing and opens the Model Download window automatically. You must select a model precision before downloading.

首次启动时，SXSEditor 检测到推理模型文件缺失，会自动打开模型下载窗口。下载前必须选择模型精度。

Precision精度 Best For适用场景 Disk Size磁盘占用
FP32 (recommended)FP32（推荐） Discrete GPU (4GB+ VRAM)独立显卡（4GB+ 显存） ~3 GB约 3 GB
FP16FP16 Integrated GPU / moderate VRAM集成显卡 / 显存适中 ~1.5 GB约 1.5 GB
INT8INT8 Integrated GPU / low VRAM (<2GB)集成显卡 / 低显存（<2GB） ~1 GB约 1 GB
INT8-NPUINT8-NPU NPU hardwareNPU 硬件 ~1 GB约 1 GB
  1. Select your desired precision. If unsure, choose FP32 for the best quality.选择所需精度。如不确定，请选 FP32 以获得最佳音质。
  2. (Optional) Click Change to choose a different download directory. The default location does not require admin privileges.（可选）点击 Change 选择其他下载目录。默认位置无需管理员权限。
  3. Click Start Download.点击 Start Download。
  4. Wait for all 9 ONNX model files to download. The window shows per-file and overall progress.等待 9 个 ONNX 模型文件全部下载完成。窗口会显示每个文件和总进度。
  5. When all files show "Complete", click Close. The main window appears.当所有文件显示 "Complete" 时，点击 Close。主窗口随即出现。
💡

The download uses chunked parallel transfer (up to 16 concurrent connections). If a download fails, click Start Download again — completed files are skipped automatically.

下载采用分片并行传输（最多 16 个并发连接）。如果下载失败，再次点击 Start Download 即可——已完成的文件会自动跳过。

Different precisions can coexist — you can switch later in Settings without re-downloading. Each precision has its own subdirectory.

不同精度可共存——之后可在设置中切换而无需重新下载。每种精度有独立的子目录。

Step 3: Create Your First Singer 第 3 步：创建第一个歌手

A singer is a voice profile built from a reference audio clip. The SVS model uses the singer's voice characteristics to synthesize new singing.

歌手 是基于参考音频片段构建的声音配置文件。SVS 模型会利用歌手的语音特征来合成新的歌声。

  1. In the main window, click the + button in the singer panel (left side).在主窗口中，点击歌手面板（左侧）的 + 按钮。
  2. In the dialog, click Open Singer Creator.在弹出的对话框中，点击 Open Singer Creator。
  3. Enter a Singer Name (e.g., "MyVoice") and choose an avatar color.输入 Singer Name（如 "MyVoice"），并选择头像颜色。
  4. Upload a pure vocal WAV file (max 30 seconds). If your file exceeds 30 seconds, a trim dialog opens — drag the selection to choose a 30-second clip, then click Confirm Trim.上传 纯人声 WAV 文件（最长 30 秒）。如果文件超过 30 秒，会自动弹出裁剪对话框——拖动选区选择 30 秒片段，然后点击 Confirm Trim。
  5. Click Start Audio Preprocessing, extract F0 and MIDI, then type the lyrics for each note. Click Save. MIDI extraction uses FCPE by default — the new recommended tool for fast, accurate F0-to-note conversion. See Settings to switch tools or adjust FCPE parameters.点击 Start Audio Preprocessing，提取 F0 和 MIDI，然后为每个音符填写歌词。点击 Save。MIDI 提取默认使用 FCPE——新的推荐工具，可快速、准确地将 F0 转换为音符。可在设置中切换工具或调整 FCPE 参数。
  6. Click Create & Save (✓). Choose a save location. The singer is saved as a .sxssinger file.点击 Create & Save（✓）。选择保存位置。歌手以 .sxssinger 文件保存。
⚠️

The reference audio must be pure vocals only — no background music, instruments, or reverb. Any non-vocal content will be learned as part of the voice and cause artifacts in synthesized output.

参考音频必须是 纯人声——不能有背景音乐、乐器或混响。任何非人声内容都会被当作声音的一部分学习，导致合成输出出现伪影。

For the full singer creation workflow, see Singer Creation.

完整的歌手创建流程请见 创建歌手。

Step 4: Add Your First Fragment 第 4 步：添加第一个片段

A fragment is a segment of music assigned to a singer. It contains MIDI notes, lyrics, and optional pitch curves.

片段 是分配给某个歌手的一段音乐。它包含 MIDI 音符、歌词和可选的音高曲线。

💡

Accompaniment tracks — you can also import accompaniment audio (WAV/MP3/FLAC/OGG/M4A/AAC) as separate tracks alongside vocal fragments. Each accompaniment track has per-track volume control and clips can be dragged to reposition. Accompaniment is mixed into the final export.

伴奏音轨——你还可以将伴奏音频（WAV/MP3/FLAC/OGG/M4A/AAC）作为独立音轨与人声片段一起导入。每个伴奏音轨有逐轨音量控制，片段可拖动重新定位。伴奏会混入最终导出。

  1. In the main window, find your singer in the left panel.在主窗口左侧面板找到你的歌手。
  2. Click the + button on the singer's row. A new fragment appears on the timeline (right side).点击该歌手行的 + 按钮。时间线（右侧）上会出现一个新片段。
  3. Drag the fragment to reposition it. Drag its edges to resize.拖动片段可重新定位。拖动其边缘可调整大小。

Step 5: Edit the Fragment (Piano Roll) 第 5 步：编辑片段（钢琴卷帘）

Double-click a fragment to open the Fragment Editor — a full piano-roll editor.

双击片段打开 Fragment Editor——一个完整的钢琴卷帘编辑器。

  1. Click on the piano roll grid to create a note. Drag while clicking to set the note length. Notes snap to the beat grid.在钢琴卷帘网格上点击创建音符。点击时拖动可设置音符长度。音符会吸附到节拍网格。
  2. Double-click a note to edit its lyric inline. For Chinese, type Chinese characters (e.g., 你好); you may append a tone digit 1–5 to force a specific tone (e.g., 你2 好3). For English, type words (e.g., hello). Pinyin text is not accepted as Chinese — only Chinese characters are.双击音符可内联编辑歌词。中文请输入汉字（如 你好）；可在汉字后加数字 1–5 强制指定声调（如 你2 好3）。英文请输入单词（如 hello）。不接受拼音作为中文输入——仅支持汉字。
  3. Move notes by dragging. Resize by dragging the right edge. Delete with Delete. Duplicate with Ctrl+D.拖动可移动音符。拖动右边缘可调整长度。按 Delete 删除。按 Ctrl+D 复制。
  4. Optionally switch to Pitch mode (press 2) to draw pitch curves, or VOL/PAN (press 3/4) for envelopes.可选：切换到 Pitch 模式（按 2）绘制音高曲线，或 VOL/PAN（按 3/4）编辑包络。
ℹ️

Edits auto-sync to the main window — there is no Save button. A 500ms debounced auto-save fires after every edit. Press Ctrl+S to force-sync immediately.

编辑会自动同步到主窗口——没有保存按钮。每次编辑后触发 500ms 防抖自动保存。按 Ctrl+S 可立即强制同步。

For the full editor reference, see Fragment Editor.

完整编辑器参考请见 片段编辑器。

Step 6: Synthesize & Listen 第 6 步：合成与试听

  1. Back in the main window, press ▶ Play in the toolbar.回到主窗口，点击工具栏的 ▶ Play。
  2. The first synthesis takes longer as the SVS pipeline initializes (loads all 9 ONNX models into GPU memory). Subsequent plays are faster.首次合成较慢，因为 SVS 管线需要初始化（将 9 个 ONNX 模型加载到 GPU 显存）。之后再次播放会更快。
  3. Use ⏸ Pause and ⏹ Stop to control playback.使用 ⏸ Pause 和 ⏹ Stop 控制播放。
💡

You can also press ▶ Play inside the Fragment Editor to preview just that fragment. This uses the "Preview Inference Parameters" from Settings (fewer diffusion steps = faster but lower quality).

你也可以在 Fragment Editor 中按 ▶ Play 仅预览当前片段。这会使用设置中的"预览推理参数"（扩散步数更少 = 更快但音质略低）。

Step 7: Export 第 7 步：导出

  1. Click 📤 Export in the main window toolbar.在主窗口工具栏点击 📤 Export。
  2. The entire project is synthesized at export quality (uses "Export Inference Parameters" — more diffusion steps for higher quality).整个项目以导出质量合成（使用"导出推理参数"——更多扩散步数以获得更高音质）。
  3. Choose a save location for the WAV file.选择 WAV 文件的保存位置。
  4. The exported file is 48 kHz, 16-bit PCM WAV by default (the model operates internally at 24 kHz and the output is linearly upsampled). Export sample rate is selectable: 24 / 44.1 / 48 / 96 kHz.导出文件默认为 48 kHz、16-bit PCM WAV（模型内部以 24 kHz 运行，输出经线性上采样）。导出采样率可选：24 / 44.1 / 48 / 96 kHz。

For details on synthesis parameters and export options, see Synthesis & Export.

合成参数和导出选项详情请见 合成与导出。

Step 8: Save Your Project 第 8 步：保存项目

What's Next? 下一步