Sample-level Sinusoidal Source Modeling

• Sample-level sinusoidal source is a signal model that synthesizes discrete signals by summing time-varying sinusoids at the audio-sample scale.
• It underpins applications like signal reconstruction, blind source separation, and neural audio decoding through optimized parameter estimation and phase tracking.
• Recent advances integrate these sources as differentiable layers in neural networks, enabling gradient-based training for precise, real-time audio synthesis.

A sample-level sinusoidal source is a generative signal model or network layer that synthesizes discrete-time signals as explicit sums of sinusoids at the audio-sample scale. In its most general form, the output at index $n$ is given by $$x[n] = \sum_{k=1}^{M} A_k[n] \cos\left(2\pi f_k[n] n / F_s + \phi_k[n]\right)$$, where the number of components $M$, frequencies $f_k$, amplitudes $A_k$, and phases $\phi_k$ may be fixed or time-varying, and $F_s$ is the sampling rate. Such representations underlie a wide spectrum of tasks including signal reconstruction from compressed domains, blind source separation, differentiable signal synthesis, and neural audio decoders.

1. Mathematical Formulation and Sample-Level Synthesis

The canonical discrete-time sinusoidal source expresses a signal as a sum of (possibly time-varying) sinusoids. In the context of audio or scientific signal processing, the model may have both fixed and adaptive parameters:

$$x[n] = \sum_{k=1}^M A_k[n] \cos(\phi_k[n]),\quad \phi_k[n+1] = \phi_k[n] + 2\pi f_k[n]/F_s$$

Here, $A_k[n]$ and $f_k[n]$ can be interpolated between analysis frames or inferred sample-wise. For framewise models, parameter estimation is typically performed at a coarser time-scale and the sample-level waveform is obtained by parameter interpolation and recursive phase updating. This framework enables synthesis that is coherent at both micro (single-sample) and macro (frame) timescales, a critical distinction from overlap-add and coarse-grained spectral approaches (Natsiou et al., 2022).

2. Estimation, Optimization, and Maximum Likelihood Approaches

Estimating sinusoidal model parameters from observed data is classically treated as a nonlinear regression task under Gaussian noise:

$$x[n] = \sum_{k=1}^{M} A_k[n] \cos\left(2\pi f_k[n] n / F_s + \phi_k[n]\right)$$0

The negative log-likelihood (dropping constants) is minimized, equivalent to minimizing the sum of squared errors:

$$x[n] = \sum_{k=1}^{M} A_k[n] \cos\left(2\pi f_k[n] n / F_s + \phi_k[n]\right)$$1

Efficient algorithms, such as the sequential extraction and least squares joint fitting of Gokcesu & Gokcesu, provide near-linear complexity solutions. Frequencies are extracted by FFT bin interpolation followed by robust local refinement using the DTFT. Amplitudes and phases are solved jointly by linear least squares after constructing a design matrix whose columns are $$x[n] = \sum_{k=1}^{M} A_k[n] \cos\left(2\pi f_k[n] n / F_s + \phi_k[n]\right)$$2 and $$x[n] = \sum_{k=1}^{M} A_k[n] \cos\left(2\pi f_k[n] n / F_s + \phi_k[n]\right)$$3 atoms at the candidate frequencies. Blind model order selection is achieved by iterative extraction with stopping criteria based on amplitude thresholds, likelihood increments, or information criteria such as AIC/BIC (Gokcesu et al., 2022).

3. Differentiable Sample-Level Sinusoidal Layers

Recent advances have enabled sample-level sinusoidal sources as trainable layers within end-to-end differentiable signal processing or neural generative models. Hayes et al. introduced a formulation using complex-exponential surrogates: $$x[n] = \sum_{k=1}^{M} A_k[n] \cos\left(2\pi f_k[n] n / F_s + \phi_k[n]\right)$$4, so that $$x[n] = \sum_{k=1}^{M} A_k[n] \cos\left(2\pi f_k[n] n / F_s + \phi_k[n]\right)$$5. Losses are defined as time-domain MSE or frequency-domain objectives.

Gradient-based training leverages Wirtinger calculus to compute partials with respect to the complex sinusoidal generator parameters. Updates are performed via first-order optimizers (e.g., Adam), frequently requiring care in initialization, regularization of $$x[n] = \sum_{k=1}^{M} A_k[n] \cos\left(2\pi f_k[n] n / F_s + \phi_k[n]\right)$$6 to constrain $$x[n] = \sum_{k=1}^{M} A_k[n] \cos\left(2\pi f_k[n] n / F_s + \phi_k[n]\right)$$7, and windowed losses to suppress local minima due to the highly non-convex error surface. This approach is critical for integrating sinusoidal source modeling in neural decoding tasks (such as DDSP) and enables differentiable, sample-level control over synthesized waveforms (Hayes et al., 2022).

4. Sinusoidal Inversion from Compressed Representations

Sinusoidal sources have proven essential for reconstructing audio from highly compressed representations, particularly the log-mel-spectrogram. The inversion pipeline proceeds by:

• Estimating the fundamental $$x[n] = \sum_{k=1}^{M} A_k[n] \cos\left(2\pi f_k[n] n / F_s + \phi_k[n]\right)$$8 per frame via YIN or related algorithms.
• Deriving the set of harmonic frequencies $$x[n] = \sum_{k=1}^{M} A_k[n] \cos\left(2\pi f_k[n] n / F_s + \phi_k[n]\right)$$9 up to the Nyquist limit.
• Estimating per-partial amplitudes $M$0 by averaging mel-band energies overlapping $M$1.
• Smoothly initializing inter-frame phases and synthesizing per-sample output via linear interpolation of amplitude and frequency, updating phase recursively.

This method achieves sample-accurate reconstruction with preserved spectral and temporal coherence and has been shown to outperform autoregressive neural vocoders (e.g., WaveNet) in STFT power MSE and spectral convergence while operating with two orders of magnitude less computation. The approach is robust for pitched monophonic sources and forms the basis for modular neural decoding layers (Natsiou et al., 2022).

5. Applications: Blind Source Separation and Audio Synthesis

Sample-level sinusoidal sources are central to blind source separation of harmonic audio mixtures. Sequentially extracting tones via log-likelihood maximization and joint least-squares fitting allows decomposition of complex, multi-tone signals without a priori knowledge of the number of constituents, provided a suitable stopping criterion. In music and speech generation, sinusoidal layers facilitate high-fidelity waveform synthesis under tight control at the level of harmonics and noise, supporting modular, interpretable, and efficient architectures. Their differentiability makes them directly suited for end-to-end optimization in neural network-based generative or denoising pipelines (Gokcesu et al., 2022, Hayes et al., 2022).

6. Numerical and Practical Considerations

Accurate sample-level modeling with sinusoidal sources requires addressing numerical and algorithmic challenges. These include:

• Robust frequency estimation in noisy or closely spaced mixtures (FFT bin interpolation plus local search).
• Avoidance of local minima in gradient-based training via good initialization (peak-picking in DFT), regularization (e.g., $M$2 decay constraints), and multiple restarts.
• Ensuring phase continuity and spectral convergence across frame boundaries by linear interpolation and explicit phase tracking.
• Balancing computational efficiency (near-linear complexity in $M$3 under mild growth conditions in $M$4 and frequency refinement) with parameter expressivity (Gokcesu et al., 2022).

A plausible implication is that further development of numerically stable, sample-level models will expand the scope and robustness of data-driven audio synthesis and signal separation frameworks.

7. Astrophysical and Spectroscopic Contexts

In astronomical time series, sample-level sinusoidal modeling aids in the classification of variable stars and the disambiguation of physical mechanisms (binarity, pulsation, rotational modulation). Šipkov et al. implemented single- and two-sine fits to TESS photometry, with radial velocity follow-up for robust distinction among binaries, pulsators, and spotted stars. This underscores the necessity of joint photometric and spectroscopic analysis for proper census and interpretation of sinusoidal variability in large-scale surveys (Šipková et al., 26 Dec 2025).

---

References:

• "Blind Source Separation for Mixture of Sinusoids with Near-Linear Computational Complexity" (Gokcesu et al., 2022)
• "A sinusoidal signal reconstruction method for the inversion of the mel-spectrogram" (Natsiou et al., 2022)
• "Sinusoidal Frequency Estimation by Gradient Descent" (Hayes et al., 2022)
• "On the origin of sinusoidal brightness variations in F to O-type stars through radial velocities" (Šipková et al., 26 Dec 2025)