Abstract
Multiple fundamental frequency (F0) estimation is a foundational task in Music Information Retrieval (MIR), yet its progress is hindered by the difficulty of obtaining accurate "ground truth" data. Traditional datasets rely on either simplistic random mixtures of samples or subjective manual annotations of real recordings, both of which introduce significant errors. This paper presents a systematic methodology for creating polyphonic music databases with verifiable ground truth. By rendering MIDI files into isolated, non-overlapping monophonic tracks and applying high-precision single-pitch estimation algorithms, the authors generate frame-accurate reference data that accounts for complex acoustic factors like instrument release times.
Introduction
In the field of MIR, identifying multiple simultaneous pitches—Multiple F0 Estimation—is often considered a "holy grail." However, evaluating these algorithms requires a reliable answer key. Currently, researchers face a trilemma: random mixtures lack musicality; real recordings require subjective and often imprecise manual annotation; and standard MIDI-synthesized audio fails to account for the "release" time of instruments, leading to timing discrepancies between the MIDI "Note Off" event and the actual end of the sound.
Methodology: Systematic Synthesis and Track Splitting
To address these limitations, the authors propose a pipeline that ensures every note in a complex polyphonic piece is automatically and accurately labeled. The core of this approach is a hierarchical track-splitting process.
Track Splitting for Clean Analysis
To eliminate ambiguity caused by overlapping notes on the same instrument, the system splits a single MIDI track into multiple sub-tracks. This ensures that for any two consecutive notes, the second note only begins after the first note's audio tail (release time) has completely faded. This mathematical accounting for $T_{release}$ eliminates the primary source of timing error found in standard synthesized datasets.
High-Fidelity Synthesis and Ground Truth Extraction
The synthesis utilizes the RWC Musical Instrument Sound Database, mapping MIDI velocities to specific instrument patches (e.g., fortissimo, mezzo-forte) for realistic dynamics. Instead of relying solely on MIDI data, the authors run the YIN algorithm—a high-precision monophonic F0 estimator—on the isolated monophonic tracks. This reduces the "multiple F0" problem to several "single F0" problems, which are scientifically solved, resulting in a "verifiable" ground truth.
Data Representation and Verification
The resulting ground truth is stored using the Sound Description Interchange Format (SDIF), a binary format designed for high-precision, frame-by-frame acoustic data. This allows for a much higher degree of temporal and frequency resolution than text-based formats like XML.
Key Insights and Results
- The Release Factor: The methodology highlights that MIDI "Note Off" events are insufficient for ground truth. By isolating tracks, the system captures the true duration of the sound.
- Musical Context: By using MIDI files of actual compositions (e.g., Mozart’s Le Nozze di Figaro), the database respects musical rules, which is essential for testing algorithms that utilize musicological priors.
- Precision and Flexibility: The method produces frame-dependent F0 tracks that discard aperiodic transients. Furthermore, the database allows for "selective polyphony," enabling researchers to render specific instrument groups to test varying levels of complexity.
Conclusion
While synthesized music still lacks the subtle expressive timing of live human performance, this systematic approach provides a significant leap forward in dataset reliability. By prioritizing verifiability over subjectivity, the authors provide a robust benchmark for the next generation of Multiple F0 estimation algorithms. Future work will focus on expanding the database to include metadata for tasks such as beat tracking and automatic orchestration.
Summary updated