머신러닝을 이용하는 음악가의 도구적 유용성에 관한 연구 : Google의 MusicLM 구현을 중심으로

국문 초록 (Abstract)

본 논문은 2023년 1월에 Google이 발표한 MusicLM(Music Language Model)을 구현하고 이를 어떻게 활용하는 것이 음악가의 작품 활동에 도움이 될지를 고찰한 연구이다. chatGPT의 생산성에 전 세계인이 매료된 것처럼 음악가뿐만 아니라 음악을 하고 싶어 하는 대중을 위한 도구가 지속적으로 제작되고 배포된다는 점은 경제적, 문화적, 사회적으로 의미하는 바가 크며 이를 연구하는 것은 음악가로서 또한 연구자로서도 기쁨이라 하겠다.
음악 산업의 성장률에 대한 기대치가 각국마다 높은 것을 보더라도 Google이 이에 관심을 갖고 관련 기술을 개발하는 것은 자연스러운 수순으로 여겨진다. 더욱이 Kontakt이나 Splice 등의 Sample 기반의 음악 제작 도구가 한창 큰 인기를 끌었던 것을 감안 한다면 이를 확장할 수 있는 가능성을 지닌 MusicLM의 활용방법이나 응용방법에 대해 알아두는 것은 여러 측면에서 유용하리라 생각한다. 또한 MusicLM이 AI(Artificial Intelligence)의 하위 분야인 머신러닝(Machine Learning)의 연구결과물이라는 점에서 다양한 분야와의 접목이 이루어질 수 있다는 것 역시 큰 장점이라 할 수 있다.
MusicLM의 가장 큰 특징은 텍스트를 입력하여 음악을 생성한다는 점이다. 이는 텍스트를 다루는 자연어 처리의 발전과 음악을 생성하고 편집하는 Music Generative 기술 간의 융합(Convergence)의 결과물이라 할 수 있다. 이러한 서로 다른 분야 간의 융합은 결국 인간감각의 총체성을 모방하고 구현하는 것을 목표로 하고 있으며 MusicLM은 이러한 방향의 극히 일부라 할 수 있는 것이다. 본 논문은 MusicLM과 관련된 기술의 세부적인 내용을 살펴보고 이를 활용한 경제적 측면, 작업의 효율성을 살펴본다. 또한 실용적이고 생산적인 내용을 포함해, 구동 원리, 음악에 대한 새로운 이해방식을 제안하고 있다. 끝으로 영감과 창의성에 관한 음악의 철학적 접근과 구체적인 방법론을 다루며 지금까지 소개된 여러 아이디어가 창작세계의 어려움을 극복할 이정표로서 기능할 것을 기대한다.

다국어 초록 (Multilingual Abstract)

This paper examines how to implement and utilize the Music Language Model (Music LM) released by Google in January 2023 will help musicians' work activities. Just as people around the world are fascinated by the productivity of chatGPT, the continuous production and distribution of tools not only for musicians but also for the public who want to do music has great economic, cultural, and social implications, and it is a pleasure as a musician and researcher to study them.
Even if expectations for the growth rate of the music industry are high in each country, it is considered a natural step for Google to be interested in it and develop related technologies. Moreover, considering that sample-based music production tools such as Kontakt and Splice were very popular, I think it will be useful in many ways to know how to use or apply MusicLM with the potential to expand it. In addition, it is a great advantage that MusicLM can be combined with various fields in that it is a research result of Machine Learning, a subfield of AI (Artificial Intelligence).
The biggest feature of MusicLM is that it generates music by entering text. This can be said to be the result of the convergence between the development of natural language processing that deals with text and the Music Generative technology that generates and edits music. These convergence between different fields eventually aims to imitate and implement the totality of human senses, and MusicLM is a small part of this direction. This paper looks at the details of the technology related to MusicLM and examines the economic aspects and work efficiency using it. It also proposes a new way of understanding the driving principle and music, including practical and productive content. Finally, it deals with music's philosophical approach to inspiration and creativity and specific methodologies, and we hope that various ideas introduced so far will serve as milestones to overcome the difficulties of the creative world.

목차 (Table of Contents)

• 표 차례 ··································································································Ⅳ
• 그림 차례 ·······························································································Ⅵ
• 1. 서론 ····································································································1
• 1.1. 연구 목적 ···················································································1
• 1.2. 연구 범위 및 방법 ······································································2
• 1.3. 선행 연구 ···················································································3
• 2. 음악과 도구의 역사 ···········································································8
• 2.1. 플루트를 연주하는 인형과 오토마타(Automata)의 역사 ·············10
• 2.2. 전자 회로의 발달과 전자악기의 출현 ········································12
• 2.3. 컴퓨터의 발달에 따른 음악 도구의 변화 ···································13
• 3. MusicLM ··························································································17
• 3.1 MusicLM의 기본 정보 ·······························································17
• 3.1.1. 소개 ··················································································17
• 3.1.2. MusicLM의 선행연구 ························································18
• 3.1.3. MusicLM의 Codes ····························································20
• 3.1.4. MusicLM의 결과물 ····························································20
• 4. MusicLM의 유용성에 관한 고찰 ·······················································23
• 4.1. 생산적 측면 ·········································································25
• 4.1.1. 경제적 측면 (An Economic Aspect) ····························25
• 4.1.1.1. 사용자의 경제적 가치를 위한 도구 ·······················25
• 4.1.1.2. 데이터세트(Dataset) 제작을 통한 저작권료 ··········27
• 4.1.1.3. 팬덤의 니즈(Needs)를 위한 전략적 도구 ··············29
• 4.1.2. 효율성 측면 ·································································32
• 4.1.2.1. 비전공자와 전공자의 음악 작업의 효율성 ·············32
• 4.1.2.2. AI와 인간 간의 헤게모니 전쟁 ·····························35
• 4.1.3. 음악 제작 기술의 이향 ·················································47
• 4.1.3.1. MusicLM의 제조 ··················································47
• 4.1.3.2. MusicLM의 편집 ··················································54
• 4.2. 인간과 기계의 학습적 측면 ··················································62
• 4.2.1. MusicLM의 학습 과정 : Operation ·····························62
• 4.2.2. 사용자에 의한 학습 : Training ····································70
• 4.2.3. 교육 도구로서의 학습 : Teaching ·······························77
• 4.3. 영감과 창의성 측면 ·····························································96
• 4.3.1. 영감의 모호성에 관한 접근과 정립(定立) ·····················96
• 4.3.2. 영감을 다루는 학문적 두 가지 갈래 ·····························98
• 4.3.3. 예측 불가의 우연 : 미지와 복잡계 ·····························100
• 4.3.3.1. 미지(未知, Unknown) ·····································100
• 4.3.3.2. 복잡계(複雜界, Complex System) ····················104
• 4.3.4. 예측 가능함의 우연 : 조합론(組合論, Combinatorics) ···112
• 4.3.5. 영감의 필수 불가결(必修 不可缺)함 : 창의성 ·············123
• 5. 결론 ·······························································································126
• 참고문헌 ····························································································· 129
• 부록 1. MusicLM Codes ·····································································132
• 부록 2. MusicLM 문장생성 결과 ·························································177
• 부록 3. 인터뷰(Interview) : 장동선 ·······················································206
• 부록 4. 인터뷰(Interview) : 정인성 ·······················································209
• ABSTRACT ··························································································212