The usage of artificial intelligence today has touched not only science but also culture. The connection between music and mathematics has always been close, because, fundamentally, the parameters of the sound organization such as rhythm, pitch, and sound ratios, refer to mathematical accuracy. But in an age of artificial intelligence, we can ask ourselves, does music still need a composer? Is technology already smart enough to create thousands of hours of music without the help of humans?
In music history, there have been some attempts to organize sounds based on certain algorithms. One of such examples could be so-called the Markov chains. The Markov chains are a random process when the future state depends on only one state of the past. This means that music is generated by memorizing the relative frequencies of the sequences and generating samples corresponding to this distribution. However, most of the attempts failed as the music sounded too mechanical. In a long run, the artificial neural networks adapted to the task of music generation can create music with a more complex structure. Template music differs from other music genres and forms because it does not require human intervention, but only examples to be followed by the artificial neural networks to create mew musical sounds.
How Does It Work?
Artificial neural networks are mathematical functions or algorithms that have been inspired by the biological neural networks. Recurrent Neural Network (RNN) is a type of artificial neural network designed to model sequences (e.g., sound sequences, texts, etc.). These networks are used to process sequence information and generate new sequences. The principle of RNN operation is based on memorized information between time series, which are especially important in modeling musical sequences when future events depend on past events, allowing the recursive neural network to reproduce a sample sound sequence with the same time series structure.
LSTM Neural Networks
However, a simple RNN network has poor long-term memory, as a result, an improved RNN architecture called LSTM (Long-short term memory) has been developed. The short-to-long memory network is a recursive network (RNN) using additional feedback to better preserve the last input signals of the period. Recursive neural networks are quite versatile, and they can be applied in music as well.
Many studies have shown that template music really has great potential to become a new way to create music. Using RNN, it is possible to create endless blues music. On the other hand, we must acknowledge the fact that the data, such as the music created by now will run out if there are no ones to expand the musical material from which RNN would be able to create music because the technology could not work without samples.
There are many doubts that an artificial neural network would be able to create a large-scale musical work, such as an opera, because such a work requires the creation of texts that could transmit certain images and a certain idea. So far, RNN are not able to understand the music theory and create templates of certain styles. For this reason, the music produced by RNN cannot be described as high quality.