musicpy

[English 中文]

Have you ever thought about writing music with codes in a very concise, human-readable syntax?

Musicpy is a python domain-specific music programming language designed to write music in a very handy syntax for musicians. It is easy to learn and write, easy to read, and incorporates a fully computerized music theory system.

Musicpy can do way more than just writing music. This package can also be used to analyze music with a music theory logic, and you can design algorithms to explore the endless possibilities of music, all with musicpy.

I've been developing many python modules and packages on my own in spare time. These python modules and packages are mainly for mathematics, statistics, games and music.

Today I wanna introduce to you a python domain-specific language developed by me that lets you write music with codes in very concise, human-readable syntax: musicpy.

This python package allows you to express notes, rhythms, dynamics and other information of a piece of music with a very concise syntax. It can generate music through music theory logic and perform advanced music theory operations. You can easily output musicpy codes into midi file format, and you can also easily input any midi files and convert to musicpy's data structures to do a lot of advanced music theory operations. The syntax of musicpy is very concise and flexible, and it makes the codes written in musicpy very human-readable, and musicpy is fully compatible with python, which means you can write python codes to interact with musicpy. Because musicpy is involved with everything in music theory, I recommend using this package after learning at least some fundamentals of music theory so you can use musicpy more clearly and satisfiedly. On the other hand, you should be able to play around with them after having a look at this wiki I wrote if you are already familiar with music theory.

Installation：

Make sure you have installed python (version >= 3.6) in your pc first. Execute the following command in the terminal to install musicpy by pip.

pip install musicpy

Importing:

Place this line at the start of the files you want to have it used in.

from musicpy import *

Because musicpy has too many features to introduce, I will just give a simple example code of music programming in musicpy:

# a nylon string guitar plays broken chords on a chord progression
guitar = (C('CM7',4, 1/4, 1/8)^2 | C('G7sus', 3, 1/4, 1/8)^2 
| C('A7sus', 3, 1/4, 1/8)^2 | C('Em7', 3, 1/4, 1/8)^2 | 
C('FM7', 3, 1/4, 1/8)^2 | C('CM7', 4, 1/4, 1/8)@1 |
C('AbM7', 3, 1/4, 1/8)^2 | C('G7sus', 3, 1/4, 1/8)^2)
play((guitar * 2)-octave, 100, instrument=25)
# or you can also write
# /(guitar * 2)-octave, 100, instrument=25
# in the IDE I write for musicpy

Introduction and Tutorial video series part 1: BV1754y197a9

Demo of writing music with musicpy: BV18z4y1r7Pk

musicpy's data sturctures, basic syntax, usage and more details are in this wiki, I'll try to make sure everything is detailed and clear, and with examples.

you can click here to download the entire wiki of musicpy I written in pdf and markdown format, which is updating continuously.

note, chord, scale are the basic classes in musicpy that builds up the base of music programming, and there are way more musical classes in musicpy.

Because of musicpy's data structure design, the note class is congruent to integers, which means that it can be used as int directly.

The chord class is the set of notes, which means that it itself can be seen as a set of integers, a vector, or even a matrix (e.g. a set of chord progressions can be seen as a combination of multiple vectors, which results in a form of matrix with lines and columns indexed)

Because of that, note, chord and scale classes can all be arithmetically used in calculation, with examples of Linear Algebra and Discrete Mathmetics. It is also possible to write an algorithm following music theory logics using musicpy's data structure, or to perform experiments on music with the help of pure mathematics logics.

Many experimental music styles nowadays, like serialism, aleatoric music, postmodern music (like minimalist music), are theoretically possible to make upon the arithmetically performable data structures provided in musicpy. Of course musicpy can be used to write any kind of classical music, jazz, or pop music.

Reasons Why I Develop This Language and Keep Working on This Project (feel free to skip this part if you are not interested)

There are two main reasons why I developed this language. First, compared to engineering files and midi files, simply storing unitized information such as notes, strength, and speed. If you can express a piece of music from the perspective of composition from the perspective of music theory, How it is achieved is more meaningful. Moreover, as long as it is not modernist atonal music, most of the music is extremely regular in music theory, which can be greatly simplified by abstracting these rules into logical sentences of music theory. (For example, a midi file with 1000 notes may actually be simplified to a few lines of code from the perspective of music theory). Second, this language was developed to allow the composition AI to compose music when it really understands music theory (rather than deep learning and feeding a lot of data). This language is also an interface. AI only needs to understand the grammar of music theory. Composing will have the same thinking as people. We can tell AI about the rules of music theory, what to do or not to do. These things can still be quantified, so musicpy can also be used as a music theory interface to communicate music between people and AI. Therefore, for example, if you want AI to learn a person’s composition style, you can also quantify the person’s style in music theory. Each style corresponds to some different music theory logic rules. These only need to be written to AI. Through musicpy, AI You can imitate that person's style. If it is AI's own original style, it is to look for possibilities in various complicated composition rules.

I am thinking that without using deep learning and neural networks to directly teach AI music theory and someone’s stylized music theory rules, then AI may be able to do better than deep learning and big data training. Because big data training is just to imitate the data itself for AI, so that in fact, AI does not really understand what composition is and what music theory is like human beings, so I want to use musicpy to teach human's music theory to AI, so that AI could truly understands music theory. In this way, the composition will not be blunt, and there will be no sense of machine and randomness. So one of my original intentions for writing musicpy is to avoid too much deep learning. But I feel that it is really difficult to abstract the music theory rules of different musicians. I will work hard to write the qwq of this algorithm. In addition, the musician himself can tell the AI how he likes to write in music theory (that is, his own unique music theory preference rules) , Then AI will imitate it well, because AI did understand music theory at that time, it is impossible to have a sense of machine and randomness in composition. At this time, what AI thinks in his mind is exactly the same as what he thinks in the mind of musicians.

AI doesn’t have to follow the rules of music theory and logic that we gave him to create. We can set a concept of “preference” to AI. AI will prefer a certain style to a certain extent when composing music, but in addition, it will have its own The unique style found in the rules of "complying with the correct music theory", so that AI can say "its own original composition style under the influence of some musicians". When this preference is 0, AI's composition will be completely based on the style found through music theory, just as a person starts to explore his own composition style after learning music theory by himself. An AI who understands music theory can easily find his own unique style to compose music. We don't even need to give him data to train, but only teach AI music theory.

So how do you teach AI music theory? Regarding music, the category of modernist music is not considered for the time being, so most of the music follows some very basic rules of music theory. The rules here refer to how to write music theory OK, and how to write music theory errors. For example, when writing harmony, four parts in the same direction are often avoided, especially when writing orchestral parts when composing music. For example, if you write a chord, if there is a minor second (or minor ninth) in the chord, it will sound more fighting. For example, when AI decides that a piece of music should be written from A major, then he should select chords from the A major scale according to the progression. It is possible to detune it appropriately, add a few subordinate chords, and finish writing the main chord. The song part may be rotated according to the fifth degree circle, or the major third/minor third, the same tonic and so on. What we need to do is to tell the AI how to write is correct when composing music, and further, how to write sounds better. AI learns music theory well, will not forget it, and it is more difficult to make mistakes, so it can write music that really belongs to AI itself. They will really understand what music is and what music theory is. Because what musicpy does is to abstract music theory into logical sentences, then every time we give AI "teaching", we express people's own music theory concepts in the language of musicpy, and then write them into the AI database. In this way, AI has truly learned music theory. Such composition AI does not require deep learning, training set, or big data. In contrast, composition AI trained by deep learning does not actually understand music theory or the concept of music. They are just taking pictures of the gourd from the massive training data. Another important point is that since things can be described with specific logic, machine learning is actually not required. If it is text recognition and image classification, which are more difficult to describe with abstract logic, that is where deep learning comes in.

summary

I started to develop musicpy in October 2019, and now I have a complete set of music theory logic syntax, and there are many composing and arranging functions as well as advanced music theory logic operations. For details, please refer to the wiki. I will continue to update musicpy's video tutorials and wiki.

I'm working on musicpy continuously and updating musicpy very frequently, more and more musical features will be added, so that musicpy can do more with music.

Thank you for your support~

Contact

qq: 2180502841
Bilibili account: Rainbow_Dreamer
email: 2180502841@qq.com

中文版介绍

[English 中文]

你们有想过用代码来写音乐吗？musicpy是一个可以让你用编程写音乐的python邻域特定语言，可以让你用非常简洁并且可读性高的语法通过乐理知识写出优美的音乐。

musicpy容易学，容易写，可读性也比较强，并且是一个完全计算机化的乐理系统。musicpy除了用来创作音乐之外，还可以从乐理层面上来创作音乐和分析音乐，并且你可以在musicpy的基础上设计乐理算法来探索音乐的可能性。

最近几个月大学的学业繁忙，但是业余时间自己开发了很多python库，内容包括数学统计，各种游戏，还有音乐等等。其实还有试着写AI方面的，但是目前还是初期进度。

今天我想先介绍一下我正在开发中的一个python库：musicpy。

这个库可以让你用非常简洁的语法，来表达一段音乐的音符，节奏，力度等等信息，可以通过乐理逻辑来生成曲子，并且进行高级的乐理操作，可以简单地输出成 midi 文件的格式，也可以很简单地输入midi文件并且转换为musicpy的数据结构进行高级乐理的操作。musicpy的语法设计非常地简洁与灵活，因此musicpy的代码的可读性比较强，并且musicpy和python完全兼容，因此你可以写python代码和musicpy进行互动。这个库里面涉及到非常多的乐理知识，所以个人推荐至少要先了解一部分乐理再来使用会比较上手。相对地，如果你是一个对乐理比较了解的人，那么看完我在 Wiki 正在写的教程之后你应该很快就上手了。

安装musicpy：

先确定电脑里有安装python，python版本最好是 >= 3.6，打开电脑的cmd然后输入

pip install musicpy

import这个库：

在你喜欢用的python的IDE里面输入

from musicpy import *

由于musicpy有太多的特性可以介绍，在这里就先写一段用musicpy语言作曲的代码示例:

# 尼龙弦吉他分解和弦演奏一个和弦进行
guitar = (C('CM7',4, 1/4, 1/8)^2 | C('G7sus', 3, 1/4, 1/8)^2 
| C('A7sus', 3, 1/4, 1/8)^2 | C('Em7', 3, 1/4, 1/8)^2 | 
C('FM7', 3, 1/4, 1/8)^2 | C('CM7', 4, 1/4, 1/8)@1 |
C('AbM7', 3, 1/4, 1/8)^2 | C('G7sus', 3, 1/4, 1/8)^2)
play((guitar * 2)-octave, 100, instrument=25)
# 或者在我为musicpy专门写的IDE里你也可以写
# /(guitar * 2)-octave, 100, instrument=25

我自己做的介绍与使用教程视频第一期：BV1754y197a9

musicpy作曲示例实际演示以及 musicpy实验作曲日常：BV18z4y1r7Pk

详细的 musicpy数据结构，基础语法以及使用教程，请看我正在写的 Wiki，我会尽量把所有的细节都讲清楚，并且提供示例代码。

我写的musicpy的wiki的pdf档和markdown档全套可以点击这里下载，正在不断地更新中

在 musicpy里面，几个基本的类是 note（音符），chord（和弦）和 scale（音阶）。这几个类是构成音乐代码的基础。除此之外，musicpy还有很多其他的乐理类型。

在 musicpy这门语言的数据结构设计中，音符类本身是等值为纯数字的，也就是完全可以作为纯数字使用。

和弦类是音符类的集合，也说明和弦类本身等值为一个全部都是数字的集合，也可以作为向量，甚至矩阵来看待（比如多个和弦的连接走向就可以看作多个向量的拼接，因此也就有了行列数，也就是矩阵的形式）。

也因此在这门语言的数据结构设计中，音符类，和弦类，音阶类都是可以进行数学运算的，比如线性代数领域的运算，离散数学领域的运算等等。也可以在这门语言的数据结构的基础上建立一整套乐理逻辑的算法，结合纯数学逻辑来进行多方面的音乐分析研究。

现代音乐领域的很多实验性质的音乐，比如序列主义，偶然音乐，后现代主义音乐（比如极简主义音乐），理论上全部都可以在这门语言的纯数字化的数据结构的基础上进行严格的创作。即使不提实验性质的音乐，这门语言也可以写任何的古典音乐，爵士音乐，流行音乐。

我开发musicpy的初衷(如果你不感兴趣的话可以跳过这部分内容)

我开发这个语言主要的初衷有两点，第一，比起工程文件和 midi 文件单纯存储音符，力度，速度等单位化的信息，如果能够按照乐理上的角度来表示一段音乐从作曲上的角度是如何实现的，那就更加有表示的意义了。而且只要不是现代主义无调性音乐，大部分的音乐都是极其具有乐理上的规律性的，这些规律抽象成乐理逻辑语句可以大大地精简化。（比如一个 midi 文件 1000 个音符，实际上按照乐理角度可能可以简化到几句代码）。第二，开发这个语言是为了让作曲 AI 能够在真正懂得乐理的情况下来作曲（而不是深度学习，喂大量的数据），这个语言也算是一个接口，AI 只要把乐理的语法搞懂了，那作曲就会拥有和人一样的思维。我们可以把乐理上的规则，做什么好不做什么好告诉 AI，这些东西还是可以量化的，所以这个乐理库也可以作为一个乐理接口，沟通人和 AI 之间的音乐。因此，比如想让 AI 学习某个人的作曲风格，那么在乐理上面也同样可以量化这个人的风格，每种风格对应着一些不同的乐理逻辑规则，这些只要写给 AI，经过我这个库，AI 就可以实现模仿那个人的风格了。如果是 AI 自己原创风格，那就是从各种复杂的作曲规则里寻找可能性。

我在想不用深度学习，神经网络这些东西，直接教给 AI 乐理和某个人的风格化的乐理规则，那么 AI 或许可以做的比深度学习大数据训练出来的更好。因为大数据训练只是给 AI 模仿数据本身而已，这样其实 AI 并没有真正地和人类自己一样理解作曲是什么，乐理是什么，所以我才想通过这个库实现把人的乐理同样教给 AI，让 AI 真正意义上地理解乐理，这样的话，作曲起来就不会生硬了，没有机器和随机的感觉了。所以我写这个库的初衷之一就是避开深度学习那一套。但是感觉抽象出不同音乐人的乐理规则确实很有难度，我会加油写好这个算法的qwq 另外其实也可以音乐人自己告诉 AI 他自己乐理上喜欢怎么写（也就是自己独特的乐理偏好规则），那么 AI 就会模仿的很到位，因为 AI 那时候确实懂得乐理了，作曲不可能会有机器感和随机感，此时 AI 脑子里想的就和音乐人脑子里想的是完全一样的东西。

AI 不必完全按照我们给他的乐理逻辑规则来创作，我们可以设置一个“偏好度”的概念给 AI，AI 在自己作曲时会有一定程度地偏好某种风格，但是除此之外会有自己在“符合正确乐理”的规则里面找到的独特的风格，这样 AI 就可以说“受到了某些音乐人的影响下自己原创的作曲风格了”。当这个偏好度为 0 时，AI 的作曲将会完全是自己通过乐理寻找到的风格，就像一个人自己学习了乐理之后，开始摸索自己的作曲风格一样。一个懂得乐理的 AI 很容易找到自己独特的风格来作曲，我们甚至都不需要给他数据来训练，而只要教给 AI 乐理就行。

那么怎么教给 AI 乐理呢？在音乐上面，暂时不考虑现代主义音乐的范畴，那么绝大部分的音乐都是遵循着一些很基本的乐理规则的。这里的规则指的是，怎么样写乐理上 OK，怎么样写犯了乐理上的错误。比如写和声的时候，四部同向往往是要避免的，尤其是在编曲时写管弦乐的部分。比如写一个和弦，如果和弦里面的音出现小二度（或者小九度）会听着比较打架。比如当 AI 自己决定一首曲子要从 A 大调开始写，那么他应该从 A 大调音阶里按照级数来选取和弦，有可能适当地离调一下，加几个副属和弦，写完主歌部分可能按照五度圈转个调，或者大三度/小三度转调，同主音大小调转调等等。我们需要做的事情就是告诉 AI 作曲的时候怎么写是正确的，更进一步的，怎么写听着比较有水平。AI 学好了乐理，不会忘记，也比较难犯错，因此可以写出真正属于 AI 自己的音乐。他们会真正懂得音乐是什么，乐理是什么。因为这个库的语言做的事情就是把乐理抽象成逻辑语句，那么我们每次给 AI “上课”，就是把人自己的乐理概念用这个库的语言来表述，然后写进 AI 的数据库里。通过这种方式，AI 真正的学习到了乐理。这样的作曲 AI，不需要深度学习，不需要训练集，不需要大数据，而与之相比，那些深度学习训练出来的作曲 AI 实际上根本就不懂乐理是什么，也没有音乐的概念，他们只是从海量的训练数据里面照葫芦画瓢而已。还有一个重点是，既然可以用具体的逻辑来描述的事情，其实是不需要机器学习的。如果是文字识别，图像分类这些比较难以用抽象的逻辑来描述的事情，那才是深度学习的用武之地。

总结

我从2019年的10月份开始开发musicpy，到现在已经有一套完整的乐理逻辑语法了，还有很多作曲编曲以及高级乐理逻辑操作的功能，详细请看wiki。musicpy的视频教程和wiki我都会持续更新。 musicpy我一直在持续更新中，不断地加入新的乐理功能，让musicpy在音乐上能做到的事情更多。

感谢大家的支持~

联系方式

qq: 2180502841
B站账号: Rainbow_Dreamer
邮箱: 2180502841@qq.com

olemb / musicpy