Hi everybody,

I’m quite sure that most indie devs would agree that finding great music for their video games is a real nightmare. Either you have to pay a lot of money to have your music composed by a decent musician, or you have to spend a zillion years searching through endless royalty-free music databases.

However, since we live in the 21st century and we have amazing technologies, I think it’s the right time to get to “music composition 2.0”. I’m a young researcher at the university, who tries to ‘teach’ computers to compose awesome music on their own. Now, I’ve decided to implement some of my academic results in order to develop a system that automatically generates customized video game music in just a few minutes.

Would you please help me improve my system by filling out this form: http://goo.gl/tg76i1? The process takes less than 3 minutes. Also, if you leave me your e-mail you'll become an alpha tester, as soon as the system goes live.

If you want to help me even more, please have a 30 minutes chat with me on Skype. Let me be clear on this: I don't want to sell anything to anyone, I just need sincere feedback from hardcore indie devs to refine my ideas! I’ll also be ready to respond to all of your questions about the project. If you are intrigued by the project and want to chat with me, send me an e-mail to velardovalerio@gmail.com.

By the way, I obviously believe that computers will become world-class composers as well as amazing painters and writers. What's your thought on this?

Best Wishes

Valerio

Interesting survey.

"Would you buy the music generated by this automatic composer, if it costed £25 per minute?" That is a rather amazing price point, especially since there are many free and low cost alternatives out there.

What research into existing systems have you done? Microsoft's DirectMusic was a fairly solid product and for about a decade tried to become viable, but never could. Dynamic soundtracks have been tried by several vendors and by several products. Generally the results range from terrible to abysmal for completely automatic music in games. Remixing systems like Parrot don't really fit well to games. There are a few custom systems for games where they start with carefully composed tracks written by expert, expensive humans, which are then mixed in a series of carefully controlled blends and transitions, these are very expensive and generally not horrible. Broad commercial viability seems a long way off.


As you are in university, what are you using for your transition basis? There have been many great research papers over the past 25 years about various generation methods and their unique pros and cons. Markov models are good but require an enormous existing corpus and is more of a remix rather than generation; genetic algorithms feel promising but so far fail because human tastes are all over the map with awful-sounding zones in between; music grammars so far require extensive expert work before anything can be generated and the results are limited to a single style; ANN generators can replay elements from emotional segments they were trained with but struggle with novelty. Attempts to blend the approaches have mixed results where they call it "musicality", it is sound that has music-based patterns, but it is not a #1 chart topping musical hit. Geraint Wiggins and some of his students have created some audio generators that some people can actually stand to listen to for a few minutes or as ambient elevator music, but not really as something people crave.

Which approach are you considering going with?

Dear frob,

thank you for your interesting and detailed reply. I have been researching the topic of automatic music generation for quite a bit now, so I’ve come to know most of the systems developed in the last 20 years. As you’ve said, most of these systems are quite unsatisfactory from an artistic point of view. However, I can list you two at least exceptions: Iamus and EMI. If you are into contemporary music, you’ll probably appreciate the music composed by Iamus (http://goo.gl/ZNzthd), Of course, it is not a AAA composer, but I would say it is a good music postgrad student! Iamus is based on a genetic algorithm with some techniques drawn from evo-devo. This system support your idea that genetic are a promising technique to generate music. EMI (Experiments in Music Intelligence) has been developed by American music professor David Cope, and it is capable of generating music in a specific style. The creative process behind EMI is quite controversial, since it is based on a kind of machine learning with some human intervention. There’s even a paper in which Wiggins harshly attacks Cope. Nevertheless, the output of EMI is quite interesting (http://goo.gl/riYiWf). However, most of the systems I’ve reviewed, focus on a specific compositional task such as creating a melody, harmonizing a choral or improvising in a real-time environment. Speaking of these systems, you’ll find some amazing algorithms. If you want an overview of the field, I suggest you to take a look at this amazing survey http://goo.gl/9R1Ppd , which describes most of the systems that deal with automatic music generation.

My academic research aims to develop a system which can compose experimental music, and can find new musical styles. To accomplish that, I use a mix of some of the techniques you’ve presented, i.e., genetic algorithms and generative grammars; along with other strategies, i.e. machine learning and multi-agent societies. To avoid the pitfalls you’ve outlined, I merge these techniques together. Following is the basic structure of the system. A ‘composer’ creates music through a genetic algorithm, with a fitness function based on some ‘musical universals’ (e.g. octave equivalence, recursive structure of music) and some ‘cultural constraints’ (e.g. tonality, cadences). The output of the ‘composer’ is sent to a society of ‘listeners’ which ‘assess’ the music and feed back to the ‘composer’, which change its generative rules accordingly.

The program which deals with video game music generation, is a simplified version of this system. It allows the user to input some musical parameters such as time-length, tempo, key, style, genre and it then composes music according to those parameters.

Hope I’ve answered your question.

b/w

Valerio

As you’ve said, most of these systems are quite unsatisfactory from an artistic point of view. However, I can list you two at least exceptions: Iamus and EMI. If you are into contemporary music, you’ll probably appreciate the music composed by Iamus ... Of course, it is not a AAA composer, but I would say it is a good music postgrad student!

I wouldn't go nearly that far.

The generative systems interest me as a programmer, but the music is still awful. I listened to the tracks you linked to, and while they do have more 'musicality', they are terrible compositions. (For my music credentials, I've played 3 instruments for 30 years now, my wife is a composer working with a local school, and I've audited several collegiate music courses on composition just to be with her as we learn new things. I'm also a semi-pro photgrapher and artist, studying those arts as well, so I can experience how different things evoke emotion.)

I still haven't heard any computer composed music that can hold a candle to even a moderately skilled composer. I have heard computers remix things well with human-guided configurations, but never heard them compose well. Those links, even though the one video had a human playing the composition, were not good music to me.

I like the research. It is a good research area that has much work to do. Just be cautious that you don't confuse mathematically interesting composition with quality artful composition.

I think that with enough information and proper rules in places, years from now software might be able to produce new and original quality music. But we are not even close to that.

Music is art. Many art classes will ask the students, "what is art?" In one class, the teacher started with what looked like a rough green 'w'. Few of the students thought it was art. It was rotated a quarter turn and suddenly appeared as the edge of a delicious-looking apple. The next image brought laughs from some Chinese students. The teacher asked before showing it for people to not tell if they knew it. When a few people suggested it was a dramatic star drawn on a dark sky, the teacher asked what it was and the Chinese students immediately answered it was the Chinese glyph for the number six.

Music is not just mathematical patterns, relationships between frequencies, harmonics, and tonal relationships. Visual art as well is not just about the patterns; some patterns are beautiful and evoke emotions, many are just noise.

Art (including both audio and visual) is about communication, usually the communication of emotion. Yes, computers can potentially make musical parts and replay what the rules specify, but there is a great deal of work needed before they produce novel evocative communications rather than somewhat interesting yet meaningless reconstructions.

For now there is no question that I would rather listen to David Garrett's reinterpretations of classics rather than a computer generated recomposition. I would also rather study the photographs of what each student selected for an alphabet photo assignment rather than just plugging each letter into Google and taking the first image. Hopefully ideas like yours will get us to the point where the algorithms can show us beauty and emotion, maybe someday, but I would not call today's constructions "good music". There is a long way to go.