AI in music production: composition, arrangement and mastering

Oblikovanje medijske budućnosti u digitalnom okruženju

Introduction: AI as a new reality in music production

Music production has always been a process that requires deep expertise, experience and creative intuition. From traditional orchestral writing to electronic music to modern digital production, each era brought new tools that transformed knowledge needed to create music [1]. However, the past five years bear witness to a fundamental transformation that differs from the previous ones: the integration of artificial intelligence in all stages of music production [2].

The AI-generated music market will reach a value of $6 billion in 2025, with forecast that it will reach over $38 billion by 2033 [3]. This figure does not contain commercial applications alone already represent a systemic change in how music is playe composes, arranges and masters. According to recent research, over 60% of musicians now uses AI tools in its production, which shows that the integration of artificial intelligence became a ubiquitous phenomenon [3].

However, this transformation is not straightforward. While AI technologies enable speed, availability and new possibilities, they simultaneously raise critical questions about nature musical creativity, authenticity and the future of musicians as a profession. This paper investigates how AI specifically affects three key stages of music production: composition (generation melody and harmony), arrangement (organization of musical elements), and mastering (final sound processing) [4].

The aim of the paper is to provide a thorough analysis of the technological possibilities of artificial intelligence in music production, explore how key neural network architectures (LSTM, Transformer, GAN), analyze research notes on the quality of AI-generated music and discuss the implications for musicians, producers and the industry as a whole. It is a special accent on understanding AI not as a replacement for shell creativity but as a tool for "collaboration" - concept that is dominant in professional practice in 2025 [2].

2026 by the authors. This article is an open access article distributed under the terms and conditions of

the Creative Commons Attribution (CC BY) license .

Technological basis: Neural network architectures in music production

LSTM networks and sequential modeling of music

By its very nature, music is a sequential phenomenon - a series of tones, rhythms and harmonics functions that take place in time [5]. This time-dependent structure requires AI system to understand how current musical elements influence future ones. Recurrent Neural Networks (RNN) and especially Long Short-Term Memory (LSTM) networks were the first to show promising results in modeling this sequential dependence [5].

The LSTM architecture allows the neural network to memorize long-term dependencies through "memory cell" mechanism. Unlike basic RNNs that quickly "forget" older sequence elements, LSTMs can memorize important musical features a few bars back. This is very important for generating coherent music because harmonic progressions, motivic elements and rhythmic patterns require longterm memory [5].

Current applications of LSTM in music production focus on MIDI generation data - standard notation for digital music display. The system is trained on a large corpus of existing music (piano concertos, pop songs, film music) and learn statistical patterns. Once trained, LSTM can generate new sequences that statistically the forms from the data collection program [4] follow. Research shows that LSTM generates music with a reasonable structure but often lacks "emotionality" and the subtlety that characterizes human composition [5].

Transformer networks and the global context

Transformer represents a revolutionary step compared to LSTMs. Instead of data process sequentially (one element at a time), Transformer can process all music elements in parallel, thanks to the "attention" mechanism [5]. This allows the model to understands global musical structures - how the whole song as a whole is organized, in what are correlation stanzas and choruses, etc. Transformer shows significantly better results in generating musical coherence and harmonic consistency than the LSTM architecture. According to recent research from In 2025, Transformer models reach near-human levels of harmonic consistency composed music, especially for standard genres such as pop and electronic music [5]. However, for more complex genres like classical music or jazz, the difference remains visible – Transformer generates music that is technically correct but lacking nuance and expressiveness [5].

Commercial AI tools like Suno, Udio and AIVA use a more advanced version of Transformer with additional layers to control genre, tempo, emotional tone and more parameters [3]. The user provides a textual description (eg "upbeat pop song about summer days"), and Transformer generates music with specific characteristics.

GAN (Generative Adversarial Networks) and audio synthesis

GAN architectures work as "two models in conflict": the generator creates a sound one file, and the discriminator tries to compare the difference between AI-generated and human of sound [4]. Through this process of competition, the generator is continuously improved. GAN shows particular potential for audio synthesis - creating realistic sounds from numerical parameters. Instead of working only with MIDI (symbolic representation music), GAN can work directly with audio waveforms, which allows for the generation of realistic instrument textures, vocal tones and even specific "curves" of playing individual

the Creative Commons Attribution (CC BY) license .

instruments [5]. This "curve" refers to An ADSR sound profile that has four stages:

• 1.    Attack — how quickly the sound reaches maximum volume (eg a drum beat is instantaneous, violin is gradual)

• 2.    Decay — fall from maximum to stable volume

• 3.    Sustain — a stable phase while the instrument is being played

• 4.    Release — how long the sound lasts after the player stops playing

Research shows that GAN-based models reach a level of harmonic consistency similar to Transformers, but with potentially greater "expressiveness" in the audio domain [5].

However, GAN is computationally more intensive and requires larger training resources than LSTM or Transformer models.

Three stages of music production through AI

Generation and composition: AI as co-author

Generation of musical composition - melody, harmonic progressions and basic structure - is the first step of music production [4]. AI systems that deal with composition usually they start with a textual description or parameters (genre, pace, emotion) and then use generative models to create a MIDI sequence [3].

The key innovation is that the AI can generate multiple versions anew and allow the producer to choose or combine different ideas. Instead of waiting for inspiration or spending an hour on writing music by hand, a producer can now get 10 different versions per minute [3]. This makes music production for all people. It helps bridge the gap, between pro music studios and people who make music at home. Now all creators can use tools. It does not matter if they are experts or not or what equipment they have.

The quality of music made this way can be different. For simpler genres (pop, electronic music, ambient music), AI generates satisfactory results [3]. For more complex genres (classical music with symphonic orchestrations, jazz with syncopation), AI-generated the music often sounds too simple or generic - technically correct but without specific character [5]. This is because these genres require a deep understanding musical conventions, stylistic nuances and emotional levels that AI is not yet complete mastered [4].

Arrangement: organization and production

Arrangement is the process of how basic melodies and harmonies are transformed into a complete one musical experience - how different instruments are used, how gradation is performed arrangements, where effects are added, etc. [4]. Traditionally, the arrangement was the most difficult task which requires experience, because it depends on the colors and characteristics of specific instruments as well as techniques.

AI systems such as AI-arrangements in DAW software (Digital Audio Workstations) or specialized tools use trained models to understand how typical arrangements for certain genre work. The system starts with a basic MIDI sequence (played by human or other AI system) and generates suggestions for adding instruments, effects, dynamic changes, etc. [4]. According to field research among producers in 2025, the arrangement is the area where AI actually has practical value. It can quickly generate the first frame - a quick version with basic instrumental layers - which the producers then fine-tune and personalize [2]. However, the final version is usually led to the collaboration of man and artificial

the Creative Commons Attribution (CC BY) license .

intelligence, where AI quickly generates options and the human chooses the best parts and adds creative ones decisions [2].

Mastering: AI in Audio Engineering

Mastering is the final stage of music production where the final mix is optimized for different platforms, speakers and devices. This requires dynamic range control (compression), frequency equalization, stereo balance and other sophisticated audio processing [6]. Traditionally, this has been the domain of specialist engineers with years of experience in this branch. AI mastering systems (such as LANDR, MasteringBOX, iZotope) analyze audio and generate recommendations or directly apply optimization using machine learning [6]. These systems were trained on the basis of thousands of professionally mastered songs, which makes it possible model to learn what a well-mastered song actually is - the so-called master [6].

However, empirical research shows an important limitation: AI mastering systems often achieve reasonable results for simple genres (pop, electronic music), but for complex genres (jazz, classical) the results are often of lower quality than the mastering that is done by an experienced sound engineer [6]. In particular, research finds that AI mastering often it causes more distortion, narrower dynamic range and more compression than is optimal [6].

This happens because AI systems often optimize songs with the goal of being loud and compressed, which is the sound characteristic of commercial pop songs. Instead, these systems should do mastering in such a way as to preserve the dynamic range (similar to classic music) which is of great importance for sound quality [6]. Mastering done by human hand is superior in preserving dynamic range, minimizing distortion and maintaining sound clarity, especially in complex genres [6].

Market and commercialization: Status in 2025

The dominant AI tools in music production

The AI music tools have really improved by this time. The leading tools are:

• •    Suno AI and Udio: Generators for complete songs with control over genre, emotion and instruments [3]. Both systems use Transformer/GAN based architecture and enable the generation of a song with a textual description.

• •    AIVA: Specialized in film scores, video games and ambient sound. Use it deep neural networks trained on professional soundtrack libraries [3].

• •    LANDR: Focuses on mastering, mixing and distribution. It uses machine learning to automatic tape optimization [6].

• •    MasteringBOX: AI-powered mastering with options to customize by genre[6].

These platforms have made music production accessible to everyone - individuals without formal experiences can now create professional sounding music [3]. This led to an explosion of new AI compositions on the Internet, which represents serious competition for musicians and producers [3].

Market and forecasts

Critical analysis of AI

The Power of AI: Where It Really Works

The AI is extremely powerful in the following scenarios:

• 1.    Fast implementation: Generating more options in a short time. For the working producer with short delivery times, this is very important [2].

• 2.    Genres with clear patterns: Pop, electronic music, ambient music - genres with relatively predictable structures – AI generates credible content [3].

• 3.    Arrangement as an initial stage: AI as an aid to quickly generate the first frame which is then improved [2].

• 4.    Mastering for simple genres: AI mastering is often satisfactory for pop or electronic music [6].

• 5.    Broad accessibility: Access to music production for people who do not have the resources or knowledge [3].

Limits of AI: Where there is no substitute for human creativity

• 1.    Complex genres: Classical music, jazz, progressive rock - genres that require stylistic sophistication – AI generates technically correct but emotionally cold music [5].

• 2.    Expressiveness: AI often generates music that sounds generic. She misses him feeling, context and specific emotion that a real composer conveys [5].

• 3.    Innovation: AI can only be inspired by existing music. New ideas, new harmonic progressions or new genres will most likely not come from an AI that is trained on existing songs [5].

• 4.    Mastering for complex genres: Experienced sound engineers are superior in the case complex music [6].

• 5.    Human intention: Music often conveys the human experience – sadness, joy, rebellion. AI which is trained on data often loses the depth of that communication [4].

Ethical and professional implications

Copyright

One of the important issues is the issue of copyright. AI systems are trained on existing music -often without explicit permission from the author [4]. This led to several lawsuits where the publishers and musicians sued AI companies [3].

The next question is: If AI composed the music, who is the author? The company that created it AI? User who asked? Both? This is not a technical problem but a legal and ethical one [4].

2026 by the authors. This article is an open access article distributed under the terms and conditions of

the Creative Commons Attribution (CC BY) license .

Dequalification of musical knowledge

If artificial intelligence can create music what does that mean for a musician who has spent years learning music theory playing an instrument and developing skills? This is a question, for musicians. One way to look at this is that music is becoming more available, to everyone, which's what I mean by democratization of music. The second is that this is a degradation of musical knowledge to triviality [3]. In practice, the trend we see is polarization: Great musicians using AI as a tool for expanding their capabilities remain relevant and valued. Musicians of mediocre ability without a specific identity they risk being replaced by AI technology [3].

Work and Economy

AI will certainly affect the economics of music production. Smaller companies can work now what used to require big production houses. But that also means it will be less musicians and engineers are needed to bring the project to an end [3].

Conclusion: The Future - Collaboration instead of replacement

This paper shows that AI has fundamentally changed music production, especially in domains where patterns are clear and generation is faster. However, the most important conclusion is that AI is not a substitute for human creativity but a new "co-producer" [2].

In practice, the musicians who are the most progressive - those who use AI to quickly generate ideas and handling routine tasks while focusing on a creative vision - they show up as the future [2]. Those who try to compete with AI as a competitor are bound to lose - because the machine can work faster, all the time and without fatigue [3].

The future of music production is a union between man and machine where man is the conductor -defining the vision, selecting the best ideas, adding emotion and context, and AI provides speed and flexibility [2].

This requires a new understanding of the role of the musician as not only a "sound maker" but as a "leader" and "interpreter of the vision". Education for musicians must include an understanding of AI tools. Ethics and law must be aligned with new technology. The industry must be reorganized around a new value – distinguishing AI quality and authenticity, especially for complex and expressive genres.

Finally, AI in music production represents a new moment where accessibility seems like dehumanization and loss of specific skills. The answer is not to say no to Artificial Intelligence or to accept it without any conditions. It is about using AI in a way that keeps what is important, in the music of creativity. Human emotion, vision, intention. And using new tools to find new possibilities. We need to think about how to use Artificial Intelligence to help us with our music of creativity. That means keeping the human emotion, vision, intention and using Artificial Intelligence to explore new ideas.