Abstract
This paper presents the ReChord Player, a concatenative synthesis engine designed to generate realistic musical accompaniments from a single recording of a performer. By utilizing chord-level audio units and a novel "Groove-Proof" mechanism, the system preserves the unique rhythmic style and micro-timing of the original musician while synthesizing new songs. The approach effectively addresses data sparsity through audio transposition and chord substitution, resulting in expressive backing tracks that maintain human-like "groove."
Introduction
Generating realistic musical accompaniment remains a significant challenge for AI. Traditional concatenative synthesis often operates at the note level, failing to capture the complex rhythmic patterns and transitions that define a musician's "groove." The authors identify two primary hurdles: sparsity of chord data in single recordings and the difficulty of distinguishing between accidental timing errors and intentional stylistic deviations. If an AI simply snaps everything to a perfect grid, the "soul" of the performance is lost; conversely, ignoring timing entirely leads to sloppy results.
Methods: Chord-Level Concatenative Synthesis
The ReChord Player treats audio chunks corresponding to specific chords as its fundamental building blocks. To generate a new accompaniment, the system employs a Viterbi-like optimization to select the most appropriate sequence of audio chunks from the source recording. It balances the "cost" of using a specific chunk based on how well it matches the target chord and how smoothly it transitions from the previous chunk.
To handle cases where the source recording lacks specific chords required for the target song, the system uses audio transposition (pitch-shifting) and chord substitution (e.g., using Cmaj7 for C). These operations are integrated into the optimization algorithm as "costs" to ensure the most natural-sounding output.
Groove Preservation
A core innovation of this work is the "Groove-Proof" engine. Human music is characterized by "micro-timing" deviations. While some are errors, others are intentional anticipations central to genres like Bossa Nova or Jazz. The system uses machine learning based on onset and harmonic features to detect these irregularities, correcting accidental errors via localized time-stretching while preserving intentional stylistic anticipations.
Results and Insights
The system was evaluated using recordings from professional Brazilian guitarists and live mobile app captures, leveraging the Lead Sheet Data Base (LSDB) containing over 10,000 Jazz and Bossa Nova standards. Results indicate that this chord-level approach significantly improves the "liveliness" and expressiveness of generated accompaniments compared to previous systems like Virtual Band. By treating the chord as the "atom" of synthesis, the engine captures the strumming patterns and resonance essential to rhythm guitar style, effectively allowing for "style transfer" of a musician's rhythmic DNA to new compositions.
Conclusion
The ReChord Player demonstrates a robust framework for real-time accompaniment generation that maintains the "feel" of a specific genre. While currently focused on Jazz and Bossa Nova, the logic is extensible to other genres. Future work aims to expand the system's genre-specific groove profiles and enhance its real-time reactivity to a live soloist's timing changes.
Summary updated