hifisampler-rs

(WIP) A Rust alternative backends of UTAU resampler hifisampler.
（仍在开发）UTAU重采样器Hifisampler的Rust替代后端。

The goal is to unofficially migrate hifisampler to Rust, just like its upstream StrayCat, in order to reduce dependencies, software size, or improve speed. This project is based on StrayCat-rs and hifisampler, and some code from StrayCrab was also referenced.
本项目的愿景是将hifisampler非官方地与其上游项目StrayCat一样迁移至rust语言中，以减少其依赖项、软件体积或提升其速度。本项目基于StrayCat-rs与hifisampler，同时还参考了部分StrayCrab的代码。

Caution

Please note that this project is not yet completed. Its performance may be inferior to the original Python version.
请注意，本项目没有开发完成。其性能可能不如原始的Python版本。

Due to certain reasons, I may not be able to allocate time to this project for a long period. Therefore, I have decided to publicly release the semi-finished code in advance, hoping to receive support from other developers.
由于一些原因，我可能很长一段时间无法分出精力在此项目上。因此我决定先行公开半成品的代码，以期待来自其他开发者的支持。

Difference 差异

During the refactoring process, this project has also made several modifications to the original workflow of hifisampler, such as merging redundant STFT calculations, utilizing real‑valued FFT, and unifying tensor layouts, etc. Due to differences in programming languages, libraries, and potential errors introduced while optimizing the computational pipeline, this project may cause discrepancies against the processing results of the original Python implementation.
在改写的过程中，本项目还对hifisampler原有的流程进行了一些改动，如合并重复的STFT计算、采用实数FFT与统一张量布局等。由于所采用的编程语言和库不同，以及在优化计算流程中可能引入的误差，本项目的处理结果可能与原来的Python实现有所不同。

Notable changes include:
主要改动包括：

Item	Python Version	Rust Version
Support model formats / 支持的模型格式	PyTorch and ONNX	Only ONNX
Configuration file / 配置文件	config.py, with vocoder‑hardcoded parameters / 含与声码器硬绑定的参数	hificonfig.ini, without vocoder‑hardcoded parameters / 不含与声码器硬绑定的参数
Features cache / 特征缓存	32 bit float npz / 32位浮点NPZ	Quantized 16 bit unsigned integer binary / 量化的16位无符号整数二进制
HNSEP cache / HNSEP缓存	32 bit float PyTorch tensor / 32位浮点PyTorch张量	Quantized 16 bit unsigned integer binary / 量化的16位无符号整数二进制

Using 使用

Warning

Since this project will ultimately be compiled into an executable file, the server-side will be renamed hifiserver-rust to distinguish it from the client(hifisampler) actually called by UTAU/OpenUTAU. To avoid such confusion in UTAU/OpenUTAU, it is not recommended to place the server-side and client-side together in the resampler folder.
由于本项目最终会编译为可执行文件，为了与UTAU/OpenUTAU实际调用的客户端(hifisampler)作区分，服务器端会被重命名为hifiserver-rust。为了避免在UTAU/OpenUTAU中混淆，不建议您将服务器端与客户端一起放入resampler文件夹中。

The client is as same as hifisampler. If you are using macOS or Linux, you can temporarily use the client of StrayCat-server.
客户端与原hifisampler的客户端一致。如果您使用macOS或者Linux，您可以暂时使用StrayCat-server的客户端。

hificonfig.ini is the server-side configuration file.
hificonfig.ini是服务器端的配置文件。

You need the following ONNX models:
您需要以下 ONNX 模型：

• pc-nsf-hifigan (or Kouon Vocoder, the pc-mini-nsf version)
• hnsep(Chinese link), a temporary version exported by myself. / 一个我导出的临时版本。

They should be located in the ./model/ folder within the same directory as the server-side, but you can also customize the model's location by modifying hificonfig.ini.
它们应位于与服务器端同目录的./model/文件夹内，但您也可以通过修改hificonfig.ini来自定义模型的位置。

How to compile

Note: By the nature of an UTAU resampler, it is only ideal to build this program in Windows.

1. Install rustup.
2. Decide whether you want to build with the icon.
   • Build with icon:
     1. Install Windows SDK.
     2. Locate rc.exe. It is usually in C:\Program Files (x86)\Windows Kits\10\bin\<version number>\x64\rc.exe
     3. Replace the location for rc.exe in the build script build.rs.
     4. Build with cargo build -r
   • Build without icon:
     1. Delete the build script build.rs.
     2. Build with cargo build -r

I highly encourage building in the other platforms as those builds can be used in OpenUtau. Build steps for Mac/Linux should be similar, just follow build without icon skipping step 1.

Supported Flags 支持的Flags

Basically same as Python version.
与Python版本基本一致。

Flags	Describe	Range	Default
g	Gender / formants 性别 / 共振峰	-600~600	0
Hb	Breath / noise 气息 / 噪波	0~500	100
Hv	Voice / harmonic 发声 / 谐波	0~150	100
HG	Vocal fry / growl 怒音 / 嘶吼	0~100	0
P1	Note level loudness normalize 音符级响度标准化	0~100	100
t	Pitch shift 音高偏移	-1200~1200	0
Ht	Tension 张力	-100~100	0
A	Amplitude 振幅	-100~100	0
G	Force regenerate cache 强制重生成缓存	bool	false
He2	Loop mode 循环模式	bool	false

You can download OpenUTAU resampler manifest file from 名無絃.
您可以下载名無絃提供的OpenUTAU重采样器配置文件。

Footnotes

1. Only effective when wave_norm is set to true in hificonfig.ini, targeting -16 LUFS.
   仅当hificonfig.ini中，wave_norm为true时有效，以 -16 LUFS 为基准。 ↩

2. Globally enabled when loop_mode is set to true in hificonfig.ini.
   当hificonfig.ini中，loop_mode为true时全局启用。 ↩