Automatic Piano Reduction Method

• Automatic piano reduction is the computational process of converting ensemble scores into piano arrangements while balancing musical fidelity and performance difficulty.
• It employs constrained probabilistic optimization and iterative inference to maximize fidelity and meet explicit playability constraints using local cost functions.
• Recent advances integrate semi-supervised transformer-based architectures and rigorous evaluation metrics to enhance reduction quality and musician acceptability.

Automatic piano reduction is the computational process of transforming a multi-instrumental (ensemble) score into a piano score, with the dual goals of preserving musical fidelity and ensuring playability within specific performance difficulty constraints. The automation of this task addresses the bottleneck in manual piano reduction for composers and performers, and has evolved rapidly through advances in statistical modeling and machine learning, culminating in systems that allow explicit control over objective musical and difficulty parameters (Nakamura et al., 2018, Wong et al., 24 Dec 2025).

1. Quantitative Performance-Difficulty Measures

Modern approaches to automatic piano reduction formulate "performance difficulty" not as an abstract or subjective label, but as a precise, local cost function derived from probabilistic generative models of piano scores. A piano reduction is represented by note pitches $p_{1:N}=(p_n)_{n=1}^N$ and onset times $t_{1:N}=(t_n)_{n=1}^N$, with a probability $P(p_{1:N})$ from a trained generative model. The local performance difficulty at time $t$ is defined as

$$D(t) = -\frac{1}{\Delta t}\,\ln P\bigl(\mathbf{p}(t)\bigr)$$

where $\mathbf{p}(t)$ is the sub-sequence of pitches in the window $[t-\Delta t/2,\,t+\Delta t/2]$. If fingerings $f_{1:N}=(f_n)$ are modeled, the joint difficulty becomes

$$D(t) = -\frac{1}{\Delta t}\,\ln P\bigl(\mathbf{p}(t),\mathbf{f}(t)\bigr)$$

Separate left-hand ($D_L$), right-hand ($D_R$), and total ($D_B=D_L+D_R$) difficulty measures are computed by restricting $\mathbf{p}(t)$ to each hand's notes. This formulation enables explicit, algorithmic control of reduction complexity in relation to performer skill and tempo (Nakamura et al., 2018).

2. Constrained Probabilistic Optimization Formulation

Piano reduction is cast as the maximization of musical fidelity under explicit difficulty constraints. Seeking a reduction $R$ of an input ensemble score $E$, the optimization aims to maximize $P(R|E)$ subject to

$$D_L(n)<\widetilde D_L,\quad D_R(n)<\widetilde D_R,\quad D_B(n)<\widetilde D_B$$

for all notes $n \in R$, where $\widetilde D$ denotes the desired difficulty threshold. Utilizing Bayes' rule $P(R|E) \propto P(R)P(E|R)$, the objective can be relaxed into a penalized log-likelihood (Lagrangian):

$$\mathcal{L}(R,\lambda_L,\lambda_R,\lambda_B) = \ln P(R) + \ln P(E|R) - \sum_{n\in R}[\lambda_L[D_L(n)-\widetilde D_L]_+ + \lambda_R[D_R(n)-\widetilde D_R]_+ + \lambda_B[D_B(n)-\widetilde D_B]_+]$$

where $[x]_+ = \max(x,0)$ encodes threshold violations (Nakamura et al., 2018).

3. Probabilistic Modeling of Musical Fidelity and Edit Processes

The modeling of musical fidelity combines two primary components:

1. Piano-score prior $P(R)$: This is instantiated as either a "no-information" uniform model, a Gaussian first-order Markov (pitch-proximity) model,

$$P(p_{1:N}) = P(p_1)\prod_{n=2}^N[\mathrm{Gauss}(p_n;p_{n-1},\sigma_p^2)+\epsilon]$$

or a fingering-aware HMM,

$$P(p_{1:N},f_{1:N}) = P(f_1)P(p_1|f_1)\prod_{n=2}^N\left[P(f_n|f_{n-1})P(p_n|p_{n-1},f_{n-1},f_n)\right]$$

2. Likelihood $P(E|R)$ ("edit model"): This describes how notes from the ensemble score $E$ are mapped to or omitted from the piano reduction $R$, introducing labels $\xi_m \in \{\mathrm{NP},L,R\}$ (not played, left-hand, right-hand) and allowing for stochastic note omission and octave shifts.

Both aspects are combined into a merged-output HMM whose state includes hand, pitch, and fingering for each token, with stochastic transitions and emission probabilities reflecting edit and pianistic constraints. This framework allows for explicit modeling of the edit operations needed to map ensemble material onto the feasible piano domain (Nakamura et al., 2018).

4. Iterative Inference and Optimization Algorithms

One-shot Viterbi inference under global constraints is typically inadequate for enforcing difficulty thresholds throughout a reduction. Therefore, a specialized iterative algorithm is employed:

• Initialize deletion-control variables $\zeta^{(1)}(m)=1$ for each ensemble note.
• Perform Viterbi decoding for a tentative reduction $R^{(1)}$.
• At each iteration, compute $D_L(n),D_R(n),D_B(n)$ for all $n \in R^{(i)}$.
• Identify and penalize ensemble notes $m$ causing difficulty violations by reducing $$\zeta^{(i+1)}(m) = \lambda \zeta^{(i)}(m),\, 0<\lambda<1$$, thereby locally increasing the likelihood of deletion.
• Re-decode on affected score segments only; iterate until all thresholds are satisfied or a maximum number of steps is reached.

This routine yields reductions that exactly meet user-specified difficulty constraints, often delivering higher average note retention and musical fidelity compared with single-pass or non-sequential models (Nakamura et al., 2018).

5. Semi-supervised and Transformer-based Architectures

To address limited availability of supervised orchestration-piano reduction pairs, recent methods employ semi-supervised deep learning, specifically utilizing BERT-style architectures adapted for symbolic music processing (Wong et al., 24 Dec 2025).

• Data and Tokenization: Pre-training is conducted on unlabeled solo-piano (GiantMIDI-Piano) and ensemble (SOD) MIDIs, utilizing MidiBERT's compound-word tokenization: each note encodes bar, sub-beat, pitch, and duration, packing fixed-length sequences.
• Two-stage pipeline:

1. Music simplification (MB-NR): A pre-trained MidiBERT encoder with a binary classification head predicts whether each orchestral note is retained or omitted.
2. Harmonization (MB-R2F): A seq2seq encoder-decoder stack (encoder: MB-NR weights; decoder: cross-attentive MidiBERT copy) reconstructs full piano textures from the sparse skeleton.

• Training regime: Strictly two-stage—first, unsupervised masked language modeling pre-training; then, fine-tuning MB-NR and MB-R2F on the small labeled LOP dataset.

Post-processing involves hand assignment (via kernel-density peak detection), duration trimming, voice limit enforcement, and smart transposition for pianistic feasibility (Wong et al., 24 Dec 2025).

6. Evaluation Metrics and Empirical Results

Objective evaluation employs the tonal similarity metric (based on pitch-class histograms in 2-beat windows) and discrimination tests involving human listeners. Professional pianist surveys directly assess musical fidelity, naturalness, and reduction quality.

Method	Tonal Similarity	Fidelity (expert)	Naturalness (expert)
MB-NR	0.39 ± 0.09	3.8	3.6
MB-R2F	0.58 ± 0.15	-	-
DBM (rule)	0.74 ± 0.07	-	-

MB-NR is experimentally preferred by professionals, passes discrimination tests (indistinguishable from human reductions), but occasionally omits melodically central material. The iterative statistical pipeline, in contrast, allows explicit trading of fidelity for difficulty and is effective at avoiding unplayable passages, especially when modeling fingerings and pitch-sequence statistics (Nakamura et al., 2018, Wong et al., 24 Dec 2025).

7. Limitations and Future Directions

Current methods face limitations such as: insufficient chordal context modeling leading to dropped important voices; neglect of inner-voice chord tones by harmonization modules; and the need for substantial post-processing engineering for playability. Prospective research includes integrating chord-prediction mechanisms into simplification heads, exploiting pseudo-labeling from large orchestral databases, and developing end-to-end difficulty-controllable reduction networks.

Further work in statistical modeling and semi-supervised architectures promises greater artistic fidelity and broader usability, enabling performers and composers to generate realistic, tailor-made reductions with constrained complexity (Nakamura et al., 2018, Wong et al., 24 Dec 2025).