Generating Nontrivial Melodies for Music as a Service (1710.02280v1)

Abstract: We present a hybrid neural network and rule-based system that generates pop music. Music produced by pure rule-based systems often sounds mechanical. Music produced by machine learning sounds better, but still lacks hierarchical temporal structure. We restore temporal hierarchy by augmenting machine learning with a temporal production grammar, which generates the music's overall structure and chord progressions. A compatible melody is then generated by a conditional variational recurrent autoencoder. The autoencoder is trained with eight-measure segments from a corpus of 10,000 MIDI files, each of which has had its melody track and chord progressions identified heuristically. The autoencoder maps melody into a multi-dimensional feature space, conditioned by the underlying chord progression. A melody is then generated by feeding a random sample from that space to the autoencoder's decoder, along with the chord progression generated by the grammar. The autoencoder can make musically plausible variations on an existing melody, suitable for recurring motifs. It can also reharmonize a melody to a new chord progression, keeping the rhythm and contour. The generated music compares favorably with that generated by other academic and commercial software designed for the music-as-a-service industry.

• The paper presents a hybrid approach that integrates conditional variational recurrent autoencoders with temporal generative grammar for context-conditioned melody generation.
• The methodology leverages a corpus of 10,000 MIDI files to train the model, ensuring melodies accurately respond to chord progressions.
• Results indicate that the generated compositions can exceed existing systems in musical quality, providing enhanced creative interaction for music-as-a-service applications.

Generating Nontrivial Melodies for Music as a Service

The paper "Generating Nontrivial Melodies for Music as a Service" explores the creation of structured and musically engaging compositions using a hybrid approach that intertwines neural networks with rule-based systems. This research takes a distinctive approach by focusing on the integration of machine learning with temporal production grammars to generate sophisticated musical pieces, particularly in the pop music genre, moving away from the often mechanical results of rule-based systems and the simplistic outputs of purely neural network-based methods.

The paper involves multiple key components that build upon existing research in the domain of automatic music generation. The authors leverage a corpus of 10,000 MIDI files to train a conditional variational recurrent autoencoder, enabling the mapping of melodies into a multi-dimensional feature space while considering the underlying chord progressions. This methodology allows for the generation of melodies that respond to chord progressions, offering the capability to produce variations of existing motifs and to adapt pre-existing melodies to new harmonic contexts.

Methodological Highlights

The research advances the framework by:

1. Conditional Variational Autoencoder: The deployment of a conditional variational recurrent autoencoder facilitates the creation of melodies that are conditioned by chord progressions, allowing for meaningful and context-aware musical variations. By incorporating a Gaussian distribution in the latent space, the model effectively learns to generate plausible melodies by sampling from this distribution.
2. Hierarchical Structure through Generative Grammar: The fusion of temporal generative grammar with a neural network architecture enables a restoration of temporal hierarchies in music composition. This contrasts with prior attempts that failed to capture this level of structural hierarchy, despite employing various neural network architectures, including LSTM networks.
3. Melody Identification and Chord Detection: The authors apply sophisticated heuristic methods to discern melodies and chord progressions from MIDI files, ensuring that the machine learning model is exposed to high-quality and relevant training data. This involves both rubric and entropy scoring to identify melody tracks and a cost-based approach for chord detection, mitigating the typical challenges associated with diverse chord voicings and modifications.

Numerical and Qualitative Results

The outcomes of the proposed system were benchmarked against existing academic and commercial solutions in the music-as-a-service industry. The results indicate that the generated music not only matches but sometimes exceeds the quality of compositions produced by these established systems. In particular, the ability to variate motifs and reharmonize melodies reflects a deeper musicality that could appeal to practitioners aiming for high levels of creative interaction and originality in generated compositions.

Implications and Future Directions

Practically, this research contributes to the domain of algorithmic composition, particularly for industrial applications in music streaming, personalized music production, and collaborative composition processes. Theoretically, it presents a novel method of harmonizing rule-based and learning-based approaches, which could serve as a reference model for future studies aiming to exploit such synergies.

Future research can focus on overcoming current limitations, such as the assumption of root position chords during chord detection, and assigning qualitative musical attributes to various dimensions of the learned representation space. Additionally, improving the generalization of the model to accommodate more diverse musical genres could expand its applicability.

In summary, by synthesizing rule-based and neural methodologies, the paper presents a comprehensive system capable of generating music compositions with considerably enriched melodic and harmonic complexity, positioned as a compelling solution for the evolving domain of music-as-a-service technologies.