Synthesis & Export 合成与导出

SXSEditor synthesizes singing voice using the SoulX-Singer diffusion model on ONNX Runtime. This page covers the synthesis pipeline, playback controls, quality settings, and WAV export.

SXSEditor 使用 SoulX-Singer 扩散模型在 ONNX Runtime 上合成歌声。本页涵盖合成管线、播放控制、质量设置和 WAV 导出。

How Synthesis Works 合成工作原理

When you press Play, SXSEditor runs the SVS pipeline:

按播放时，SXSEditor 会运行 SVS 管线：

1. Prepares input: Collects MIDI notes, lyrics, pitch curve, and singer data.准备输入：收集 MIDI 音符、歌词、音高曲线和歌手数据。
2. Processes text: Converts lyrics to phoneme sequences using language-specific processing.文本处理：使用语言特定处理将歌词转换为音素序列。
3. Encodes: Runs 5 encoder models (text, pitch, note type, F0, condition embedding).编码：运行 5 个编码器模型（文本、音高、音符类型、F0、条件嵌入）。
4. Runs diffusion: Iteratively denoises a mel spectrogram using the diffusion model.扩散：使用扩散模型迭代地对梅尔频谱图去噪。
5. Vocalizes: Converts the mel spectrogram to a waveform using the vocoder.声码器：使用声码器将梅尔频谱图转换为波形。
6. Plays audio: Outputs the waveform through the audio system.播放：通过音频系统输出波形。

Playback Controls 播放控制

• ▶ Play: Start synthesis and playback. On first play, the SVS pipeline initializes (loads models into GPU).▶ Play：开始合成与播放。首次播放时，SVS 管线会初始化（将模型加载到 GPU）。
• ⏸ Pause: Pause playback. Resume from the paused position.⏸ Pause：暂停播放。从暂停位置继续。
• ⏹ Stop: Stop playback and reset to the beginning.⏹ Stop：停止播放并重置到开头。

Fragment-Level Playback 片段级播放

In the Fragment Editor, ▶ Play synthesizes and plays only the current fragment. This is useful for quick previews while editing. It uses the "Preview Inference Parameters" from Settings (fewer diffusion steps = faster response).

在片段编辑器中，▶ Play 仅合成并播放当前片段。便于编辑时快速预览。它使用设置中的"预览推理参数"（扩散步数更少 = 响应更快）。

Preview vs Export Quality 预览与导出质量

SXSEditor uses two sets of inference parameters. Preview uses fewer steps for faster response; Export uses more steps for higher quality. Both are configurable in Settings.

SXSEditor 使用两套推理参数。预览使用较少步数以加快响应；导出使用更多步数以获得更高音质。两者均可在设置中配置。

Diffusion Sampler 扩散求解器

The sampler is the ODE solver used inside the diffusion loop. It decides how each denoising step combines model evaluations (NFE = number of function evaluations) into the mel-state update. Four solvers are available; the preview and export samplers are configured independently in Settings > Diffusion Sampler:

求解器是扩散循环中使用的 ODE 求解器，决定每个去噪步如何将模型推理（NFE = 模型评估次数）组合为 mel 状态更新。提供四种求解器；预览与导出的求解器可在 设置 > 扩散求解器 中独立配置：

• Euler (default) — 1 NFE/step. First-order explicit, midpoint time evaluation. Fastest, the original baseline.Euler（默认）— 每步 1 次推理。一阶显式，中点时间评估。最快，原始基线。
• Heun — 2 NFE/step. Second-order improved Euler (trapezoidal rule). Higher accuracy, roughly 2× inference time. Final step falls back to Euler to avoid t > 1.Heun — 每步 2 次推理。二阶改进欧拉（梯形公式）。精度更高，推理时间约 2 倍。末步退化为 Euler 以避免 t > 1 越界。
• Extrapolated Euler — 1 NFE/step. Velocity-extrapolation heuristic inspired by STORK (ICLR 2026). Reuses the previous step's velocity for linear extrapolation; not the full stabilized RK formulation. Best when the velocity field changes smoothly; falls back to Euler when extrapolation is unsafe.Extrapolated Euler — 每步 1 次推理。受 STORK（ICLR 2026）启发的速度外推启发式。复用上一步速度做线性外推；非论文完整 stabilized RK 格式。速度场缓变时效果最佳；外推不安全时退化为 Euler。
• STORK-2 — 1 NFE/step. Paper-faithful Stabilized Taylor Orthogonal Runge-Kutta (Tan et al., ICLR 2026, arXiv:2505.24210). Runge-Kutta-Gegenbauer 2nd-order recurrence with 8 sub-stages and Taylor-expansion virtual NFE. Designed for stiff ODEs with extended stability region (~2s² = 128×); higher per-step algebraic cost than Euler.STORK-2 — 每步 1 次推理。论文原版 Stabilized Taylor Orthogonal Runge-Kutta（Tan et al., ICLR 2026, arXiv:2505.24210）。Runge-Kutta-Gegenbauer 二阶递推 + 8 个 sub-stage + Taylor 展开 virtual NFE。专为刚性 ODE 设计，稳定性域扩展约 2s²=128 倍；每步代数运算开销高于 Euler。

Diffusion Steps 扩散步数

The diffusion model iteratively denoises a mel spectrogram. The number of steps controls the trade-off between quality and speed:

扩散模型会迭代地对梅尔频谱图去噪。步数控制质量与速度的权衡：

• More steps = higher quality, slower inference. Recommended for export: 32 steps.更多步数 = 更高音质，推理更慢。导出推荐：32 步。
• Fewer steps = faster inference, lower quality. Recommended for preview: 8–16 steps.更少步数 = 推理更快，音质略低。预览推荐：8–16 步。
• Range: 4–64 steps.范围：4–64 步。

CFG Strength CFG 引导强度

Classifier-Free Guidance (CFG) strength controls how closely the output adheres to the input conditions (notes, lyrics, pitch):

无分类器引导（CFG）强度控制输出与输入条件（音符、歌词、音高）的贴合程度：

• Higher values = more aligned with conditions, but may over-strengthen. Recommended: 1.0–5.0.较高值 = 更贴合条件，但可能过度强化。推荐：1.0–5.0。
• Set to 0 to skip unconditional prediction entirely. This roughly doubles synthesis speed since only the conditional pass runs.设为 0 可完全跳过无条件预测。由于只运行条件分支，合成速度约提升一倍。
• Range: 0–10.范围：0–10。

CFG Rescale CFG Rescale 系数

The CFG rescale coefficient mitigates over-guidance artifacts caused by high CFG strength. Recommended: 0.5–0.9. Range: 0–1.

CFG rescale 系数用于缓解高 CFG 强度造成的过度引导伪影。推荐：0.5–0.9。范围：0–1。

Exporting to WAV 导出为 WAV

Exporting a Single Fragment 导出单个片段

In the Fragment Editor, click 💿 Export:

在片段编辑器中，点击 💿 Export：

1. The Export dialog opens, pre-filled with the export-quality parameters from Settings. You can override them for this export only: Sampler (Euler / Heun / Extrapolated Euler / STORK-2), Diffusion Steps, CFG Strength, CFG Rescale, Auto Shift, plus advanced options — Dynamic Thresholding (enable/disable, percentile 0.9–0.999) and CFG Schedule (constant / linear / cosine / custom) (the global Settings values are not changed).弹出 导出对话框，预填设置中的导出质量参数。可仅为本次导出覆盖：求解器（Euler / Heun / Extrapolated Euler / STORK-2）、扩散步数、CFG 强度、CFG Rescale、Auto Shift 以及高级选项——动态阈值（启用/禁用，百分位 0.9–0.999）和 CFG 调度（恒定 / 线性 / 余弦 / 自定义）（不会修改全局设置）。
2. After confirming, the fragment is synthesized using the chosen parameters.确认后，片段按所选参数合成。
3. A file save dialog appears. Choose a location and filename.弹出文件保存对话框。选择位置和文件名。
4. The WAV file is saved at the selected export sample rate (default 48 kHz, selectable: 24 / 44.1 / 48 / 96 kHz, 16-bit PCM).WAV 文件按所选导出采样率保存（默认 48 kHz，可选：24 / 44.1 / 48 / 96 kHz，16-bit PCM）。

Exporting the Entire Project 导出整个项目

In the main window, click 📤 Export:

在主窗口中，点击 📤 Export：

1. The Export dialog opens (same options as single-fragment export: Sampler, Diffusion Steps, CFG, Auto Shift, Dynamic Thresholding, CFG Schedule).弹出 导出对话框（选项与单片段导出相同：求解器、扩散步数、CFG、Auto Shift、动态阈值、CFG 调度）。
2. All fragments are synthesized sequentially using the chosen parameters.所有片段按所选参数顺序合成。
3. Fragments are mixed together according to their timeline positions.片段根据时间线位置混合在一起。
4. A file save dialog appears. Choose a save location.弹出文件保存对话框。选择保存位置。
5. The mixed WAV file is saved at the selected export sample rate (default 48 kHz, selectable: 24 / 44.1 / 48 / 96 kHz, 16-bit PCM). Accompaniment audio tracks are mixed into the export.混合后的 WAV 文件按所选导出采样率保存（默认 48 kHz，可选：24 / 44.1 / 48 / 96 kHz，16-bit PCM）。伴奏音轨会混入导出。

Export Progress 导出进度

During export, a progress indicator shows:

导出过程中，进度指示器会显示：

• Current fragment being processed.正在处理的当前片段。
• Overall progress percentage.总进度百分比。
• Status messages (preparing, synthesizing, encoding WAV, saving).状态消息（准备中、合成中、编码 WAV、保存中）。

Export-Time Options 导出时选项

The Export dialog exposes several options that can be overridden per export without changing global Settings:

导出对话框提供多项可仅为本次导出覆盖的选项，不会修改全局设置：

• Sampler selection — choose Euler / Heun / Extrapolated Euler / STORK-2 for the export diffusion pass (see Diffusion Sampler in Settings).求解器选择——为导出扩散过程选择 Euler / Heun / Extrapolated Euler / STORK-2（见设置中的扩散求解器）。
• Dynamic thresholding — enable/disable and adjust the percentile (0.9–0.999). When enabled, extreme values in the diffusion latent are clipped at the specified percentile, reducing artifacts from over-sharp predictions.动态阈值——启用/禁用并调整百分位（0.9–0.999）。启用后，扩散潜空间中的极端值按指定百分位裁剪，减少过度尖锐预测导致的伪影。
• CFG schedule — choose how CFG strength varies across diffusion steps: constant (same strength throughout), linear (ramp from high to low), cosine (smooth cosine curve), or custom (user-defined per-step values).CFG 调度——选择 CFG 强度在扩散步间的变化方式：恒定（全程相同强度）、线性（从高到低渐变）、余弦（平滑余弦曲线）或 自定义（用户逐步指定）。

Exporting LRC Lyrics 导出 LRC 歌词

A new 📄 Export LRC button in the toolbar allows exporting the project lyrics as a timed LRC file. The LRC file contains timestamped lyric lines synchronized with the project timeline, compatible with standard LRC players and karaoke software.

工具栏新增 📄 Export LRC 按钮可将项目歌词导出为带时间戳的 LRC 文件。LRC 文件包含与项目时间线同步的带时间戳歌词行，兼容标准 LRC 播放器和卡拉 OK 软件。

Multi-Segment Synthesis 多段合成

Long audio is synthesized in segments. SXSEditor aligns F0 across segments using absolute-time alignment and applies per-segment autoShift f0Shift for wide-range songs, so multi-segment prompt mel frame counts stay consistent with the single-segment path. The vocoder processes audio in chunks (default 1008 frames, overlap 8 frames) with crossfade stitching.

长音频会分段合成。SXSEditor 使用绝对时间对齐跨段的 F0，并针对宽音域片段应用 per-segment autoShift f0Shift，使多段 prompt mel 帧数与单段路径一致。声码器按分片处理音频（默认 1008 帧，重叠 8 帧），交叉淡入淡出拼接。

Output Format 输出格式

Performance Tips 性能优化建议

• Use a discrete GPU if available. CPU inference is significantly slower.如有 独立显卡 请使用。CPU 推理明显较慢。
• Reduce diffusion steps in Settings > Preview Inference Parameters. Default is 16; try 8 for faster preview.在设置 > 预览推理参数中 减少扩散步数。默认 16；可尝试 8 以加快预览。
• Set CFG Strength to 0 to skip unconditional prediction, roughly doubling speed.将 CFG Strength 设为 0 可跳过无条件预测，速度约提升一倍。
• Update GPU drivers to the latest version for best DirectML performance.更新显卡驱动 到最新版本以获得最佳 DirectML 性能。
• Use FP16 precision instead of FP32 for lower VRAM usage and faster inference on most GPUs.在大多数显卡上 使用 FP16 精度 代替 FP32 可降低显存占用并加快推理。
• The first synthesis is always slower — subsequent syntheses reuse the loaded pipeline.首次合成总是较慢——之后的合成会复用已加载的管线。