This article comes from “ music preamble ” (ID: nakedmusic), author: Liu branching
Recently, the American digital research institute Space150 conducted an interesting experiment: based on artificial intelligence technology, imitating the vocal and musical style of the well-known rapper Travis Scott, and making a rap robot “Travis Bott”.
The purpose of this experiment is to see what AI can continue to create. In fact, “Travis Bott” really wrote a song “Jack Park Canny Dope Man”, and wrote the lyrics and melody by himself. At the same time, Space150 also used AI-based human image synthesis technology “Deepfake” to shoot the music video for this song. To be honest, unlike the previous AI songs, this AI song has almost reached the end of the real human hearing after continuing to learn from real people. Foreign netizens commented below the MV: “better than real trvis. (It is better than real people) ” “Pretty amazing, this is only the beginning. (Awesome This is just the beginning) “, and even began to worry that AI would enslave human beings, but they would still buy tickets to watch.
In principle, Space150 uses additional neural network technology (Additional Neural Network) to create melody and percussion accompaniment, and then add Travis Scott’s lyrics Enter “Text Generator Model (Text Generator Model) “, two weeks later, AI “Travis Bott” started to create the rhyme of the lyrics (rhymes) . Judging from the effect, Travis Bott imitates Travis Scott almost to the point where it is false and completely integrates the most prominent characteristics of Travis Scott’s works and the charm of the characters, so that he can be teased to join Spotify’s rap hit song list “Rap Caviar “. At the same time, the project also further validated the progress of artificial neural network technology, which is helpful to explore the value of AI in music in the future.
It is undeniable that AI has gradually been embedded in our daily lives. In the context of the new era of Internet + and Industrial Manufacturing 4.0, it is a general trend that AI composition with communication, network and human-computer interaction functions covers education science, art performance and entertainment services. In the face of the outstanding performance of AI music, let us also think: Whether musicians will co-exist with AI music, will they encounter AlphaGo-type crushing?
How do I clone Travis Scott?
In fact, AI composition (Algorithmic Composition, also known as “algorithmic composition”) is not unusual, and it is not difficult to copy Travis Scott.
As early as 2016, Sony Computer Science Labs (Computer Science Laboratories, Sony CSL for short) Researchers Hajeris and Pachter have developed a” DeepBach (Deep Bach) “neural network. They used Deeper Bach’s 352 works to train DeepBach and created 2503 hymns.
The first AI virtual composer to officially gain world status is AIVA launched by Aiva Technologies, a startup born in 2016. (Artificial Intelligence Virtual Artist ) . Its creative direction is mainly classical music, film and television soundtrack, and other types of works, such as rock music and pop music, have gradually developed to the present. As a virtual musician, it has been legally registered by the French and Luxembourg Author Rights Association (SACEM) and has its own signed copyright. In the field of AI, the work of copying the music style of one or more musicians may already be in progress.
At present, whether it is DeepBach, AIVA or Travis Bott, behind AI composition is a kind of deep learning based on artificial neural network. (Deep Learning) technology. In this kind of deep learning, programmers must build a multi-layer “neural network” and program it separately in a multi-layer structure so that they can process information between various input and output points.
Source: 2017 · Pineapple Science Award
DeepBach’s input is 362 works of Bach, AIVA’s input information is a large database of works of classical composers represented by Bach, Beethoven, Mozart, etc., while Travis Bott’s input is Travis Scott’s Works, voices and sound effects. After the data is entered, the artificial neural network will find the rules that exist among many input works, and then form an understanding of the music style. But this music style is not the final product. Its main purpose is to predict. The AI program will continue to run with its prediction of music style, and the next verification data set will be encountered in front. This data set will tell it whether the prediction is correct, and the correct and wrong feedback will be remembered by the AI. In the continuous high-speed learning, the AI’s predictive ability will become stronger and stronger. Style of music, and then you can write your own music. The breakthrough of the AI creator “Travis Bott” lies in that it is not only the input of Travis Scott’s works, but also the input of human voice and sound effects. The input and output of text and sound have taken a step forward in deep learning.
Deep learning from the “I am AI” series of short documentaries seems to be based on a simple model of the neural structure of the human brain, but in a way it can already “think” like humans. This also enables AI to understand and shape highly abstract models in data, such as models in melody, or features of human faces. But from the evolution of artificial intelligence music, artificial neural network is only one of the main technologies of AI composition. Compared with other algorithms, it has its advantages and disadvantages. In terms of advantages, the ability of artificial neural networks to excel at other algorithms is that it has self-learning capabilities, associative storage capabilities, and the ability to find optimal solutions at high speed.
Source: 2017 · Pineapple Science Award, Interpretation of Artificial Intelligence Themes
But its disadvantages are also obvious: 1. The famous “black box”The problem means that you do n’t know how the neural network will produce results, and you do n’t know why it will produce such results; 2. Unlike cognition, composition is a higher-level intelligent activity; 3. Time-consuming and labor-intensive; 4 Data 饕餮, compared with traditional machine learning algorithms, it requires more data; 5. The cost of computing power is more expensive.
In practice, even the most advanced deep learning algorithms require weeks to fully train a successful deep neural network. At present, there is no optimal solution in the main technology of AI composition, and most of them use the hybrid algorithm (Hybrid Algorithm) .
How to avoid the copyright risk of AI composition?
At the same time, the overall lack of AI composition has also become apparent. As mentioned earlier, AI composition is essentially big data and cloud computing. The process of AI music generation is that the machine summarizes and extracts the features that match the input from the programmer based on the elements or patterns entered by the programmer, and then These features extract various data elements for new combinations or extensions. There must be a question in this: How does this huge database distinguish which data is copyrighted? What is public data? How does the database builder protect the rights and interests of copyrighted data? How can the subject using the database not to infringe?
Obviously, the current AI composition is still unable to complete or to complete this task to a certain extent. Most of the circumvention of copyright comes from the intention of programmers. In 2017, Aiva Technologies’ explanation of AIVA’s choice to focus on classical music also responded to the programmer’s deliberate design of the copyright of AI composition: the classical music database used to train Aiva does not involve copyright issues because the copyrights have expired.
For Travis Bott at the beginning, in the study of Travis Scott, the sampling of the work library and character images must also be authorized by Travis Scott first, but how can the works produced after his study avoid the Travis Scott What about plagiarism? This situation is also one of the reasons for the uneven quality of AI composition in the current market. To some extent, plagiarism may be difficult to avoid. Checking tool (Plagiarism Checker) and the checking scale are particularly important here, but in terms of current practice, human musicians Song plagiarism judgment standards are still seeking to unify, so what about AI composition?
And even if the AI composition finally produces a purely original work that does not involve any infringement after hard work, he / she will face the problem of copyright certification . According to the definition of copyright in China’s “Copyright Law”: Copyright is the right granted to civil subjects to works and related objects by copyright law. Among them, the civil subject refers to citizens, legal persons or unincorporated organizations. AI cannot be recognized in the identity of the subject, and the acquisition and renunciation of rights have become more complicated. If infringement disputes arise, it will be difficult to resolve. For example, Microsoft Xiaobing’s independent collection of poems “Sunshine Lost Glass Window”, once the work was published, a lot of piracy and a lot of irregular references appeared. This kind of infringement in the usual sense is due to the lack of lack of legal provisions, so that the ownership of copyright is unclear, and infringement is left to the discretion.
But it is worth mentioning that compared to domestic blanks, the relaxation and recognition of AI works abroad has become a norm. Britain, South Africa, and New Zealand are among the first countries to explicitly recognize AI copyright. Although the United States, Japan, and Australia have not clearly stipulated in statutory law, they have tried to varying degrees in judicial practice. This is why the United States has not recognized AI works in statute law, but has won cases in judicial practice. However, because China is a statutory law country, case law is not a formal source of law, and cannot form a judge-made law with the common law system (or common law system) Judicial practice, so clearly defining AI works from the system is the most fundamental.
It is undeniable that it may take some time to obtain a wide range of recognition due to the differences in the level of AI and legal operations in many countries. Of course, it is also relatively simple to make a coincidence. Adding the name of a human artist to an AI-generated work can break through this confusion. On September 7, 2018, the practice of AIVA’s pure music album “Ai (Vol.3 from artificial composer Aiva) ” was: AIVA, but each piece will be marked “feat. Aiva Sinfonietta Orchestra, Brad Frey”, indicating that the music supervisor’s contribution in the “performance”, the team members can commercialize the work.
In general, copying Travis Scott is not difficult for AI, but it is not a day’s work to handle copyright disputes and further improve AI technology.
The commercial exploration of AI music
AI music is undoubtedly a long-standing, but booming industry in recent years. In 1974, the advent of the Rader system was the real beginning of an AI composition system. Different from AI in the present sense, it uses the rules that can be used in AI, so that the machine can make tradeoffs based on the rules of melody and harmony generation, and choose the appropriate proportion of notes and harmony. Since then, with the continuous deepening of research on music generation systems, Snobol systems that can complete automatic bass harmony generation, andChoral system with Bach-style harmony (Ebciogln product, expert system) . In 1993, the Musact system using artificial neural network learning mode for harmony generation, and the Harmonet system based on the combination of artificial neural network and “limited satisfaction technology”, which can generate baroque harmony according to melody. These are the originators of modern AI composition systems and are landmark.
The development of contemporary AI composition systems has mostly started with Google’s Magenta. Magenta is an artificial intelligence technology that Google open sourced at the end of 2015 and uses TensorFlow machine engine learning. This project aims to develop AI technology to create music and other art forms. The main sub-projects are NSynth Super, Onsets and Frames and MusicVAE. Since then, various AI systems and products have developed rapidly. Among them, the more representative program developments are: Amper Music application used in Taryn Southern’s album “I Am AI” in 2017, and Flow Machines (Sony product) tools, and a deep neural network MuseNet developed by OpenAI in 2019 for generating musical works.
At present, more mature AI music companies abroad include Popgun in Los Angeles, Jukedeck and AI Musical in London, Humtap in San Francisco, Melodrive in Berlin, and Groov at Google headquarters in Mountain View, in addition to Google, Sony, and Amper Music. .A, AIVA in Luxembourg, OpenAI, a non-profit research company, and self-proclaimed “The first full-service record company based on artificial intelligence music discovery “Snafu Records” and others. Among them, Jukedeck was acquired by BYTE in July 2019.
In China, AI music also has many industry practices. In addition to Baidu, Tencent, Ali, Netease Cloud and other music platforms all have AI music layouts to varying degrees, universities and large and medium-sized enterprises have gradually joined the education and research and development of AI music. For example, Ping An Technology Co., Ltd., which seems to be incompatible with music, has successively cooperated with universities such as the Central University for Nationalities and the Sichuan Conservatory of Music, and won the intelligent competition in the 2018 AI Composition International Challenge held by the Swiss Federal Institute of Technology in EPFL. The first AI World Composition Competition in the composition field.
The AI music technology developed by the artificial intelligence creativity team of the Microsoft (Asia) Internet Engineering Institute has been able to create content based on multiple musical elements such as chords, rhythms, and melody crossing, and compose, compose, arrange, and sing. Waiting for a number of musical creations, it is equivalent to a complete band. Today, this technology has been repeatedly verified in CCTV and various provinces and cities’ variety shows, and has successfully achieved commercialization and industrialization output. In May 2018, Microsoft announced that the company’s artificial intelligence Xiaobing had mastered the lyrics creation and composition ability.
In addition, in April 2018, the music AI creation assistant “Little Hi” released by Hi Flip House has already createdSeveral albums, in addition to lyric and composing, also have the function of “knowledge”. The “Whaling” APP launched on IOS and Android in February and March 2019, respectively, is a music application that can make ordinary people’s online chorus possible.
It is not difficult to find that the use of AI in the field of music has become a major focus of cultural industries in various countries. While developing rapidly, there are certain difficulties. Of course, it mainly revolves around algorithms and copyright. However, with the improvement of the overall technical level of AI and the increase of users’ requirements for the intelligence of the composition system, the use of AI in the music field is gradually out of the predicament, and the domestic development trend is gradually in line with international standards. First, in terms of algorithm technology, hybrid algorithms and personalized intelligent music customization are still mainstream. On the one hand, because various algorithms have their own advantages and disadvantages in the use of artificial intelligence composition, the current style and genre of musical works composed by artificial intelligence are relatively single and not audible. In the hybrid algorithm composition, various algorithms will strengthen the strengths and avoid the weaknesses. These problems can be effectively solved. On the other hand, because AI composition rules are extracted from big data, it is prolific but can easily cause problems with high homogeneity of songs. However, personalized intelligent music customization is premised on the personal preferences of the audience. The works produced by big data and algorithms are also more original due to individual differences.
Second, in terms of copyright, subject to the insurmountable legal predicament, the shift of AI technology to cooperation with human musicians will be the most direct means to break copyright in the short term; at the same time, human musicians will also profit from it. The inspiration for human creativity and the inspiration of musicians will become more prominent. It has been reported that man-machine cooperation is 20 times faster than human musicians. To a certain extent, AI composition has advantages that human collaboration can hardly reach in terms of improving the work efficiency of musicians and reducing the communication costs between musicians and producers.
In September 2018, Yao Music, the chief scientist of Ali Music, said at the Ali Music Forum: “I think that any artist always needs inspiration when his creativity is exhausted. Music created by AI may not be the entire song They all sound good, but there is a short section in the middle that matches the emotions of these artists, and artists can use them as a starting point for inspiration and translate this inspiration into their own works. I think this is very helpful for them. ” With the gradual deepening of AI technology in deep learning, the gradual mastery of human emotions, and the gradual improvement of the definition of computer works and subjects by law, the status quo of AI as an auxiliary tool for human musicians may not last long. After all, technology and law are not static.
Conclusion
From the use of AI in streaming media for intelligent recommendation to guide listeners’ music tastes, to the creation of AI-based composers based on AI once again to disrupt the music industry, people have mixed feelings about the development of AI. On the one hand, the addition of AI can make the music industry more complete and make the operation of this industry more efficient; on the other hand, as a human-made machine, the sales and quality of AI composition may make many musicians ashamed. In the long run, the relationship between AI and human musicians and radio DJs may not be the other way around. Just like the confrontation between digital music and vinyl, the decline of vinyl is obvious, but its value is still recognized by the public. , Even sought after by a small number of people. In other words, the advancement of technology and the comprehensive advancement of the industry will most likely make AI music a standard configuration for music creation. Of course, people have higher requirements on the originality and aesthetics of human musicians in music.
But whether it is AI music or music created by human beings, the core of music products from the birth of music to now is still providing services. This core will not change, and the relationship between people and music will not be changed. . In the final analysis, artificial intelligence is still derived from human wisdom. Rather than letting musicians lose their jobs or encountering AlphaGo-style crushing, it is better to say that the industry has changed due to technology. In terms of choice of works or music services, listeners also More diverse options.
References
“ARTIFICIAL INTELLIGENCE MADE A SONG IN THE STYLE OF TRAVIS SCOTT. IT SOUNDS UNNERVINGLY LIKE TRAVIS SCOTT.” “Music Business Worldwide”, February 16, 2020
How do artists view the future of virtual reality? “” SIZE Trendy Life “, February 16, 2020
How did AIVA, the first formal AI composer in the world, make music? “” Lake World “, March 17, 2017
《What is artificial neural network (ANN)》