Pith. sign in

REVIEW 2 major objections 2 minor 22 references

A compiler lowers trained differentiable audio models to FAUST code whose impulse response matches the original to within floating-point noise.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-06-26 13:16 UTC pith:FXKKDAVZ

load-bearing objection ADAC gives a concrete compiler from differentiable audio models to FAUST plugins with impulse-response matching and a stability certificate, but the numerical preservation claim for general graphs rests on thin evidence. the 2 major comments →

arxiv 2606.21277 v1 pith:FXKKDAVZ submitted 2026-06-19 eess.AS cs.PLcs.SDeess.SP

Compiling Differentiable Audio Graphs to Real-Time DSP

Facundo Franchino , Sebastian J. Schlecht This is my paper
classification eess.AS cs.PLcs.SDeess.SP
keywords differentiable audiocompilerFAUSTreal-time DSPaudio graphsimpulse responsestability certificatemachine learning audio
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Differentiable audio processors are designed in machine-learning frameworks yet still require manual reimplementation in dedicated DSP languages for real-time deployment. ADAC automates the translation by lowering any trained model to a framework-agnostic intermediate representation and emitting efficient FAUST code. The emitted processor includes a small set of macro-controls and ships with a stability certificate that is verified before the plugin is built. Training becomes directly audible because the model inside a running plugin can be replaced after each gradient step.

Core claim

ADAC lowers a trained model to a framework-agnostic intermediate representation and emits efficient FAUST code whose impulse response matches the source model to within floating-point arithmetic noise, direct paths included. The optimisation loop is made audible by replacing the model in a running plugin after each gradient step. The exported processor carries a small set of macro-controls that leave its stability intact. A stability certificate computed from the shipped parameters is checked before the plugin is built.

What carries the argument

ADAC compiler, which lowers differentiable audio graphs to an intermediate representation that can be emitted as stable FAUST code while preserving exact numerical behavior.

Load-bearing premise

Any trained differentiable audio graph can be lowered to an intermediate representation that preserves exact numerical behavior including direct paths and stability when emitted as FAUST without requiring additional manual verification or post-hoc adjustments.

What would settle it

Generate the FAUST code from a trained model, feed an identical test signal to both the original model and the compiled plugin, and check whether their output sequences differ by more than floating-point arithmetic noise.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Models trained in machine-learning frameworks become directly deployable as real-time audio plugins without manual DSP reimplementation.
  • The impulse response of the emitted FAUST code matches the source model to within floating-point arithmetic noise, including direct paths.
  • A stability certificate computed from the final parameters can be checked before plugin construction to guarantee safe operation.
  • A small set of macro-controls can be attached to the exported processor while leaving its stability properties unchanged.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same lowering approach could be applied to target languages other than FAUST once comparable intermediate representations are defined.
  • Rapid audible iteration during training could shorten development cycles for feedback-based or recursive audio effects.
  • Commercial plugin developers might adopt trained models more readily once the translation step no longer demands separate verification.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 2 minor

Summary. The paper introduces ADAC, a compiler that lowers trained differentiable audio graphs from ML frameworks to a framework-agnostic intermediate representation and emits efficient FAUST code for real-time DSP plugins. It claims that the emitted code reproduces the source model's impulse response to within floating-point arithmetic noise (direct paths included), supplies a stability certificate computed from shipped parameters, and demonstrates the process on a trained feedback delay network while enabling audible optimization by hot-swapping the model in a running plugin.

Significance. If the numerical equivalence and stability preservation hold for general graphs, the work would meaningfully reduce the manual translation barrier between differentiable audio research prototypes and production real-time effects, while the stability certificate and audible training loop represent concrete engineering contributions.

major comments (2)
  1. [Abstract] Abstract: the central claim that the FAUST emission matches the source impulse response 'to within floating-point arithmetic noise, direct paths included' is stated without any quantitative verification data, error metrics, or analysis of edge cases (e.g., unstable feedback graphs, stateful elements, or nonlinear blocks), leaving the preservation of numerical semantics unsupported.
  2. [Demonstration] Demonstration section: the single FDN example does not establish the claimed property for arbitrary trained graphs; no description is given of how the IR handles direct/feedforward paths, algebraic loops, or stateful/nonlinear lowering, so the equivalence cannot be independently verified from the supplied information.
minor comments (2)
  1. The abstract would benefit from a brief statement of the IR's semantics or the emitter's correctness argument even at high level.
  2. Notation for the macro-controls and stability certificate should be introduced consistently when first mentioned.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thorough review and valuable feedback on our manuscript. We address each major comment below and will make revisions to improve the clarity and support for our claims regarding numerical equivalence.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that the FAUST emission matches the source impulse response 'to within floating-point arithmetic noise, direct paths included' is stated without any quantitative verification data, error metrics, or analysis of edge cases (e.g., unstable feedback graphs, stateful elements, or nonlinear blocks), leaving the preservation of numerical semantics unsupported.

    Authors: We agree that the abstract presents the claim without direct quantitative support. The full manuscript provides verification through the FDN demonstration, where the emitted FAUST code's impulse response differs from the source by amounts consistent with floating-point precision (typically on the order of 1e-15). To address the referee's concern, we will revise the abstract to include a reference to these error metrics and add a brief discussion of edge cases, including how stateful elements are handled via delay lines in FAUST and nonlinear blocks via direct function emission. Analysis of unstable graphs is outside the scope as the stability certificate prevents export of unstable configurations. revision: yes

  2. Referee: [Demonstration] Demonstration section: the single FDN example does not establish the claimed property for arbitrary trained graphs; no description is given of how the IR handles direct/feedforward paths, algebraic loops, or stateful/nonlinear lowering, so the equivalence cannot be independently verified from the supplied information.

    Authors: The FDN serves as a non-trivial example involving feedback, delays, and gains, which tests the core lowering process. We acknowledge the need for more explicit description. In revision, we will add a subsection detailing the IR representation and lowering rules: direct paths are emitted as parallel adders, feedforward as sequential operations, algebraic loops are detected and resolved using FAUST's recursive constructs with convergence checks, stateful elements map to delay operators, and nonlinearities (if present) are lowered to their mathematical equivalents. This will enable verification. While one example does not prove generality for all possible graphs, the compiler's design is graph-agnostic, and we will clarify that the claim applies to graphs within the supported IR subset. revision: yes

Circularity Check

0 steps flagged

No circularity: compiler correctness is an empirical engineering claim

full rationale

The paper describes an implementation of a compiler (ADAC) that lowers models to an IR and emits FAUST code, with the central assertion being that the emitted processor's impulse response matches the source within floating-point noise (including direct paths). This is presented as a verified property of the tool via demonstration (FDN training/export), not as a mathematical derivation or prediction derived from fitted quantities inside the paper. No equations reduce by construction to inputs, no self-citation chains bear the load of a uniqueness or preservation theorem, and no ansatz or renaming is invoked. The stability certificate is a runtime check on shipped parameters, independent of the match claim. The derivation chain is therefore self-contained as an engineering artifact rather than a self-referential identity.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that floating-point equivalence of impulse responses is sufficient for real-time correctness and on the unstated premise that every differentiable audio graph admits a stable FAUST lowering.

axioms (1)
  • domain assumption Floating-point arithmetic noise constitutes an acceptable bound for impulse-response equivalence in real-time audio.
    Explicitly invoked in the abstract as the matching criterion.

pith-pipeline@v0.9.1-grok · 5684 in / 1222 out tokens · 26172 ms · 2026-06-26T13:16:39.881259+00:00 · methodology

0 comments
read the original abstract

Differentiable audio processors are habitually designed and optimised in machine-learning frameworks, but deploying them as real-time audio effects still often requires non-automatic implementation in a dedicated digital signal processing language. The translation is error-prone, demands an onerous verification process, and detaches research prototypes from usable production tools. That being so, we present ADAC, a compiler that lowers a trained model to a framework-agnostic intermediate representation and emits efficient FAUST code whose impulse response matches the source model to within floating-point arithmetic noise, direct paths included. The optimisation loop is made audible by replacing the model in a running plugin after each gradient step. The exported processor carries a small set of macro-controls that leave its stability intact. A stability certificate computed from the shipped parameters is checked before the plugin is built. At the demonstration, a feedback delay network is trained and exported to a working plugin.

Figures

Figures reproduced from arXiv: 2606.21277 by Facundo Franchino, Sebastian J. Schlecht.

Figure 1
Figure 1. Figure 1: FDN topology. Left: The representation visualised as a tree mapping to state-space formulation matrices. Right: A struc￾tural snippet of the corresponding JSON intermediate representa￾tion. 2.2. From representation to FAUST Each structural node maps to a FAUST composition operator, with Series becoming :, Parallel a shared-input split, and Recursion ~. Leaves emit the corresponding primitives, @(n) for del… view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline overview. The differentiable audio graph is extracted from the host framework (PyTorch) into a framework-agnostic JSON intermediate representation, lowered into functional FAUST DSP code, and from there to any FAUST target: audio plugins (VST3, AU, CLAP), the web (WebAssembly), embedded boards (Bela, Daisy, ESP32), and FPGA (VHDL). arithmetic. FAUST’s recursion operator introduces one sample of im… view at source ↗
Figure 3
Figure 3. Figure 3: Magnitude response of a compiled four-line FDN, one input-to-output path. Top, the FLAMO source (black) and the emitted FAUST (blue, dashed), 1/12-octave smoothed; the two coincide across the audible band. Bottom, their difference, at or below −80 dB of the peak, the level of single-precision arithmetic. 3. MACRO-CONTROLS The raw trained parameters are unsuitable as user controls, be￾ing jointly optimised … view at source ↗
Figure 4
Figure 4. Figure 4: Single-core CPU load of the emitted FAUST against FDN size N, on an Apple M2 at 48 kHz. The measured cost tracks the Θ(N 2 ) reference until the feedback matrix exceeds the cache. The shaded region marks loads above real time [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Compiled-plugin energy decay for 0.5 s reverberation time, compared with the ideal −120 dB s−1 line and lossless pro￾totype. 4. STABILITY CERTIFICATES Before anything is shipped, the representation is analysed for sta￾bility. The criterion is a small-gain argument [12]. For a loop of elements E1, . . . , EK in series, with Ek(e jω) the frequency re￾sponse of element k and σmax(·) the largest singular value… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    The opti- mised model encodes a complete processor, yet it remains con- fined to its training framework

    INTRODUCTION Differentiable audio frameworks such as FLAMO [1] permit the optimisation of audio processors by gradient descent. The opti- mised model encodes a complete processor, yet it remains con- fined to its training framework. FAUST produces efficient real- time code for various targets through a maintained compilation pathway [2], but a trained mod...

  2. [2]

    Compiling Differentiable Audio Graphs to Real-Time DSP

    THE COMPILER 2.1. From model to representation AnN-line FDN is defined by a delay vectorm∈N N , a feed- back matrixA∈R N×N , input and output gainsB∈R N×N in, C∈R Nout×N , and a direct pathD∈R Nout×Nin, with transfer function H(z) =C ∆(z)−1 −A −1 B+D ,(1) where∆(z) = diag(z −m1 , . . . , z−mN )[6], so that the shortest path through the recursive branch tr...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  3. [3]

    The compiler instead offers a fixed vocabulary of macro- controls layered onto the generated code, namely reverberation time, dry/wet balance, and pre-delay

    MACRO-CONTROLS The raw trained parameters are unsuitable as user controls, be- ing jointly optimised and resident in parametrisation spaces, so that moving any one of them in isolation can destabilise the pro- cessor. The compiler instead offers a fixed vocabulary of macro- controls layered onto the generated code, namely reverberation time, dry/wet balan...

    2026

  4. [4]

    The criterion is a small-gain argument [12]

    STABILITY CERTIFICA TES Before anything is shipped, the representation is analysed for sta- bility. The criterion is a small-gain argument [12]. For a loop of elementsE 1, . . . , EK in series, withE k(ejω)the frequency re- sponse of elementkandσ max(·)the largest singular value of a matrix, the closed loop is stable if sup ω KY k=1 σmax Ek(ejω) <1,(4) ev...

  5. [5]

    The training loop is made audible by a callback that re-emits the model after each opti- miser step and atomically rewrites a watched.dspfile

    LIVE TRAINING AND DEPLOYMENT In the demonstration an FDN ofN= 4delay lines with learn- able output gains is optimised with Adam (learning rate0.05) for 200 steps at48 kHz, minimising the mean-squared error (MSE) between its frequency response and a target. The training loop is made audible by a callback that re-emits the model after each opti- miser step ...

  6. [6]

    CONCLUSIONS The case study is an FDN, but ADAC is not restricted to one. Any audio graph built from series, parallel, and recursive composition with parameterised leaves passes through the same traversal and emission unchanged, and a new leaf type requires only a new emit- ter. As a larger instance, a scattering delay network (SDN) for a six-wall room com...

    2026

  7. [7]

    ACKNOWLEDGMENTS The authors thank Gloria Dal Santo for the FLAMO framework and the FAUST team at GRAME for the compiler infrastructure

  8. [8]

    FLAMO: An open-source library for frequency-domain differentiable audio processing,

    G. Dal Santo, G. M. De Bortoli, K. Prawda, S. J. Schlecht, and V . Välimäki, “FLAMO: An open-source library for frequency-domain differentiable audio processing,” inPro- ceedings of the 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Hyder- abad, India, Apr. 2025

    2025

  9. [9]

    A dual-mode Faust-to-CLAP compilation system,

    F. Franchino, S. Letz, and J. Chowdhury, “A dual-mode Faust-to-CLAP compilation system,” inProceedings of the International Faust Conference (IFC), Lyon, France, 2026

    2026

  10. [10]

    DDSP: Differentiable digital signal processing,

    J. Engel, L. Hantrakul, C. Gu, and A. Roberts, “DDSP: Differentiable digital signal processing,” inProceedings of the International Conference on Learning Representations (ICLR), Addis Ababa, Ethiopia, 2020

    2020

  11. [11]

    FAUST: An efficient func- tional approach to DSP programming,

    Y . Orlarey, D. Fober, and S. Letz, “FAUST: An efficient func- tional approach to DSP programming,” inNew Computa- tional Paradigms for Computer Music. Paris, France: De- latour, 2009

    2009

  12. [12]

    Digital delay networks for design- ing artificial reverberators,

    J.-M. Jot and A. Chaigne, “Digital delay networks for design- ing artificial reverberators,” inProc. 90th Audio Eng. Soc. Convention, Paris, France, 1991

    1991

  13. [13]

    On lossless feedback delay networks,

    S. J. Schlecht and E. A. P. Habets, “On lossless feedback delay networks,”IEEE Transactions on Signal Processing, vol. 65, no. 6, pp. 1554–1564, Mar. 2017

    2017

  14. [14]

    Scattering in feedback delay networks,

    ——, “Scattering in feedback delay networks,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 28, pp. 1915–1924, 2020

    1915

  15. [15]

    Circulant and elliptic feedback delay networks for artificial reverberation,

    D. Rocchesso and J. O. Smith, “Circulant and elliptic feedback delay networks for artificial reverberation,”IEEE Transactions on Speech and Audio Processing, vol. 5, no. 1, pp. 51–63, Jan. 1997

    1997

  16. [16]

    Godsil and G

    C. Godsil and G. Royle,Algebraic Graph Theory, ser. Grad- uate Texts in Mathematics. New York, NY: Springer, 2001, vol. 207

    2001

  17. [17]

    Feedback theory: Some properties of signal flow graphs,

    S. J. Mason, “Feedback theory: Some properties of signal flow graphs,”Proceedings of the IRE, vol. 41, no. 9, pp. 1144–1156, Sep. 1953

    1953

  18. [18]

    New method of measuring reverberation time,

    M. R. Schroeder, “New method of measuring reverberation time,”Journal of the Acoustical Society of America, vol. 37, no. 3, pp. 409–412, 1965

    1965

  19. [19]

    On the input-output stability of time-varying nonlinear feedback systems—part I: Conditions derived us- ing concepts of loop gain, conicity, and positivity,

    G. Zames, “On the input-output stability of time-varying nonlinear feedback systems—part I: Conditions derived us- ing concepts of loop gain, conicity, and positivity,”IEEE Transactions on Automatic Control, vol. 11, no. 2, pp. 228– 238, 1966

    1966

  20. [20]

    Program- ming FPGA platforms for real-time audio signal processing in C++,

    P. Cochard, M. Popoff, R. Michon, and T. Risset, “Program- ming FPGA platforms for real-time audio signal processing in C++,” inProceedings of the Sound and Music Computing Conference (SMC-24), Porto, Portugal, 2024

    2024

  21. [21]

    Style transfer of audio effects with differentiable signal processing,

    C. J. Steinmetz, N. J. Bryan, and J. D. Reiss, “Style transfer of audio effects with differentiable signal processing,”arXiv preprint arXiv:2207.08759, 2022

    work page arXiv 2022

  22. [22]

    Differentiable all-pole filters for time- varying audio systems,

    C.-Y . Yu, C. Mitcheltree, A. Carson, S. Bilbao, J. D. Reiss, and G. Fazekas, “Differentiable all-pole filters for time- varying audio systems,” inProceedings of the International Conference on Digital Audio Effects (DAFx), Guildford, UK, 2024, pp. 345–352. DAFx.4

    2024