Introduction
In music theory, keys and chords are fundamentally interdependent; a musical key provides the structural context that renders a sequence of chords coherent. Despite this relationship, automated Music Information Retrieval (MIR) systems often treat chord transcription and key extraction as isolated tasks. This research addresses two primary challenges in the field: the chronic shortage of large-scale, human-annotated musical datasets and the inherent harmonic ambiguity of chords when analyzed without global context.
Methodology
The proposed system employs a unified framework inspired by isolated word recognition in speech processing.
Key-Specific HMMs
The architecture consists of 24 parallel Hidden Markov Models (HMMs), representing the 12 major and 12 minor keys. For a given audio input, the system calculates the likelihood across all models. The model with the highest probability identifies the global key, while the Viterbi path through that specific model provides the chord sequence.
Supervised Training via MIDI
To circumvent the lack of labeled audio data, the authors utilized 1,046 MIDI files. Automated harmonic analysis of MIDI data provided precise ground-truth labels, which were then synthesized into approximately 55 hours of audio for supervised training. This approach demonstrates that models trained on synthetic data can generalize effectively to real-world acoustic recordings.
Tonal Centroid Features
Instead of standard 12-dimensional chroma features, the system utilizes a 6-dimensional Tonal Centroid vector. This feature projects audio into a space defined by the Circle of Fifths and the Circle of Thirds, effectively capturing essential harmonic relationships while reducing spectral noise.
Key Insights
The core innovation lies in the use of key-specific transition matrices. Chord transitions are highly dependent on the tonal center; for instance, a G major chord is significantly more likely to resolve to C major in the key of C major than in other keys. By incorporating these context-specific probabilities, the HMM can effectively disambiguate noisy audio signals.
Results and Evaluation
The system was evaluated against high-quality ground-truth annotations of early Beatles recordings.
- Chord Recognition: The key-specific models achieved a frame-rate accuracy of 74.36%, surpassing a universal, key-independent baseline (72.75%).
- Key Estimation: The system correctly identified the key for 84.62% of the test tracks (22 out of 26).
- Performance: The unified approach outperformed contemporary state-of-the-art models, particularly in scenarios where beat-synchronous processing was not employed.
Limitations and Future Work
While successful, the current model is limited by its training data (primarily rock music) and a simplified chord vocabulary restricted to major and minor triads. Future research will focus on developing genre-specific models and incorporating higher-order HMMs to capture more complex musical phrases and a wider variety of chord types, such as diminished or seventh chords.
Summary updated