MacOS Mojave 安装 AI 東北きりたん 东北切蒲英 NEUTRINO 教程

参考【使い方あり】AIシンガー「AIきりたん」が凄すぎる。【まとめ】

• 东北切蒲英（日语：東北きりたん）是SSS合同会社发表的东北支援角色，是以日本秋田县北部的鹿角市的地方美食切蒲英（日语：“きりたんぽ”）为原型而设定的角色。同时具有UTAU、 NEUTRINO、CeVIO AI歌唱音源和VOICEROID语音声源。
• 2020年2月22日神秘的软件工程师SHACHI，发布了东北切蒲英的音源演示歌曲，同时也公开配布了基于深度学习的歌声合成引擎NEUTRINO，只要交给她乐谱就可以很熟练的唱起来。¹
  

下载NEUTRINO与KIRITAN音源

前往NEUTRINO官网，点击Start Now来到官方的谷歌云盘，下载以下两个文件，解压备用

NEUTRINO-macOS-v1.0.0.zip
歌声ライブラリ(Singer Library)/東北きりたん（NEUTRINO-Library）-v1.0.0.zip

导入并使用KIRITAN作为默认音源

将解压出的KIRITAN文件夹放置到NEUTRINO/model，然后编辑NEUTRINO/Run.sh，将ModelDir改为KIRITAN

# NEUTRINO
ModelDir=KIRITAN
StyleShift=0

安装gcc

该软件使用了gcc编译，而mac默认只有clang，因此还需要安装gcc。下面命令应该默认安装gcc11版本，如果安装了其他版本，请强制指定版本为11。

brew install gcc

修复动态链接库错误

NEUTRINO/bin/musicXMLtoLabel的动态链接库指向有误，需要额外修复

cd bin
# 原链接
otool -L musicXMLtoLabel
musicXMLtoLabel:
	/Users/user213944/.homebrew/opt/gcc/lib/gcc/11/libstdc++.6.dylib (compatibility version 7.0.0, current version 7.29.0)
	/usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 1311.100.3)
	/Users/user213944/.homebrew/opt/gcc/lib/gcc/11/libgcc_s.1.dylib (compatibility version 1.0.0, current version 1.0.0)

# 进行修改
install_name_tool -change /Users/user213944/.homebrew/opt/gcc/lib/gcc/11/libstdc++.6.dylib /usr/local/opt/gcc/lib/gcc/11/libstdc++.6.dylib musicXMLtoLabel
install_name_tool -change /Users/user213944/.homebrew/opt/gcc/lib/gcc/11/libgcc_s.1.dylib /usr/local/opt/gcc/lib/gcc/11/libgcc_s.1.dylib musicXMLtoLabel

# 再次查看
otool -L musicXMLtoLabel 
bin/musicXMLtoLabel:
	/usr/local/opt/gcc/lib/gcc/11/libstdc++.6.dylib (compatibility version 7.0.0, current version 7.29.0)
	/usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 1311.100.3)
	/usr/local/opt/gcc/lib/gcc/11/libgcc_s.1.dylib (compatibility version 1.0.0, current version 1.0.0)

合成测试

使用sample1进行合成测试。修改Run.sh的BASENAME为sample1

# Project settings
BASENAME=sample1
NumThreads=4

执行Run.sh，观察输出

sistine:NEUTRINO rumia$ ./Run.sh 
03:15 : start MusicXMLtoLabel
    Convert MusicXML to label -> score/musicxml/sample1.musicxml
    output full label -> score/label/full/sample1.lab
    output mono label -> score/label/mono/sample1.lab

03:15 : start NEUTRINO
    NEUTRINO - NEURAL SINGING SYNTHESIZER (Electron v1.0.0-Stable)

    Linguistic feature (duration)       : 1 [msec]
    Load timing model                   : 241 [msec]
        -> Load completed. 
        -> Tohoku Kiritan - NEUTRINO Singer Character Library (v1.0.0-Stable-Timing model)
    Predict timing feature              : 409 [msec]
    Linguistic feature (acoustic)       : 712 [msec]
    Load acoustic model                 : 726 [msec]
        -> Load completed. 
        -> Tohoku Kiritan - NEUTRINO Singer Character Library (v1.0.0-Stable-Acoustic model)
    Predict acoustic features           : 2051 [msec]
    Finish                              : 66688 [msec]
    Generation rate                     : 0.647793 [gen/sec]

    -- File and Parameter information --
    label length     	: 129 [line]
    wav length       	: 43.2 [sec] 
    frame period     	: 5 [frame]
    full_label       	: score/label/full/sample1.lab 
    timing_label     	: score/label/timing/sample1.lab 
    output f0        	: ./output/sample1.f0 
    output mgc       	: ./output/sample1.mgc 
    output bap       	: ./output/sample1.bap 
    model directory  	: ./model/KIRITAN/ 
    stat_timing      	: ./model/KIRITAN/stats_timing.bin 
    model timing     	: ./model/KIRITAN/model_timing.bin 
    stat_acoustic    	: ./model/KIRITAN/stats_acoustic.bin 
    model acoustic   	: ./model/KIRITAN/model_acoustic.bin 
    timing flag      	: 0 
    random flag      	: 0 
    acoustic flag    	: 0 
    number of threads	: 4 
    style shift      	: 0 
    ------------

04:21 : start WORLD
    WORLD - NEUTRINO Edition (v1.0.0-Stable)

    Load Acoustic features   : 1 [msec]
    Decode Acoustic features : 7 [msec]
    Synthesis                : 186 [msec]
    Finish                   : 1518 [msec]
    Generation rate          : 28.4585  [gen/sec]
    -- File and Parameter information --
    wav Length               : 43.2 [sec]
    sampling rate            : 48000 [Hz]
    sampling bit             : 16 [bit]
    pitch shift              : 1 
    formant shift            : 1 
    number of parallel       : 4 [thread]
    hi-speed synthesis       : 0 
    realtime synthesis       : 0 
    smooth pitch             : 0 
    smooth formant           : 0 
    enhance breathiness      : 0 
    -------------------

04:23.2N : start NSF
    NSF_IO - Neural Source Filter (I/O) (v1.0.0-Stable)

    Linguistic feature (duration)       : 5 [msec]
    Linguistic feature (acoustic)       : 279 [msec]
    Separate feature                    : 289 [msec]
    Synthesis (NSF)                     : 344 [msec]
    Error: Failed to run NSF. Please check log (NSF/NSF.log). 
05:15 : END

执行到最后一步NSF时出现了问题，手动替换变量尝试重新执行

sistine:NEUTRINO rumia$ ./bin/NSF_IO score/label/full/sample1.lab score/label/timing/sample1.lab output/sample1.f0 output/sample1.mgc output/sample1.bap KIRITAN output/sample1_nsf.wav -t
    NSF_IO - Neural Source Filter (I/O) (v1.0.0-Stable)

    Linguistic feature (duration)       : 0 [msec]
    Linguistic feature (acoustic)       : 239 [msec]
    Separate feature                    : 251 [msec]
    Synthesis (NSF)                     : 310 [msec]
    Write wav                           : 160424 [msec]
    Finish                              : 160666 [msec]
    Generation rate                     : 0.268881 [gen/sec]
    -- File and Parameter information --
    label length     	: 129 [line]
    wav length       	: 43.2 [sec] 
    frame period     	: 5 [frame]
    full_label       	: score/label/full/sample1.lab 
    timing_label     	: score/label/timing/sample1.lab 
    input f0         	: output/sample1.f0 
    input mgc        	: output/sample1.mgc 
    input bap        	: output/sample1.bap 
    nsf directory    	: NSF/ 
    output f0        	: NSF/output/f0/ 
    output mgc       	: NSF/output/mgc/ 
    output bap       	: NSF/output/bap/ 
    output list      	: NSF/output/wav.list 
    output wav       	: output/sample1_nsf.wav 
    ------------

成功。现在合成的最终结果已经输出到了output/sample1_nsf.wav。效果非常好，接近人声。

编辑乐谱

可以使用MuseScore编辑乐谱，然后将乐谱放到score，同样修改Run.sh的BASENAME，重新执行即可。

---

1. 东北切蒲英 - 萌娘百科 ↩︎