Improving Neural Pitch Estimation with SWIPE Kernels

David Marttila; Joshua D. Reiss

REVIEW 3 major objections 6 minor 33 references

Improving Neural Pitch Estimation with SWIPE Kernels

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper shows that SWIPE, a 2008 sawtooth-inspired pitch estimator, is more accurate than the state-of-the-art self-supervised neural detector PESTO on two music benchmarks, and that SWIPE scores as an audio frontend let a 647-parameter…

desk verdict A transparent, useful empirical paper whose small margins make its 'original form beats PESTO' headline slightly too strong, but the SWIPE-frontend result is solid. read the letter →

arxiv 2507.11233 v1 pith:DKFZ425T submitted 2025-07-15 cs.SD eess.AS

David Marttila , Joshua D. Reiss This is my paper

classification cs.SDeess.AS

keywords pitchestimationSWIPEself-supervisedlearningaudiofrontendToeplitzlayerfundamentalfrequencyDSPspectralkernels

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

SWIPE, a hand-crafted pitch estimator from 2008 that matches the input spectrum against sawtooth-wave kernels, turns out to be far more accurate than previous neural-network papers reported, and in the configuration recommended for music it outperforms PESTO, the current state of the art in self-supervised neural pitch detection. The authors show that feeding SWIPE scores to a neural network as an audio frontend lets a tiny 647-parameter encoder match or exceed PESTO's accuracy while using roughly forty times fewer parameters, and that the same frontend improves cross-domain generalization and noise robustness in supervised training. If these claims hold, classical DSP features deserve a serious second look as both standalone estimators and learned-frontend replacements, and the reported baselines in the self-supervised pitch literature are understated.

What carries the argument

The machinery is the SWIPE kernel: for each candidate fundamental frequency, a spectral template with positive lobes at integer harmonics, decaying like a sawtooth spectrum, and negative valleys halfway between harmonics. The SWIPE' variant removes the non-prime-harmonic lobes to suppress octave errors, and the score is a normalized inner product between the kernel and the square-root magnitude spectrum, computed at 1024 mel-spaced frequencies. This kernel explicitly encodes periodicity, so when the scores are used as a frontend the neural network only has to reweight them and refine the peak location, which is why a single Toeplitz layer suffices. A detail that carries much of the argument is the choice of mel-scale sampling rather than ERB, which the original SWIPE paper recommended for speech; the authors use mel for music.

What would settle it

Run the released SWIPE code with ERB sampling instead of mel on MIR-1K and MDB-stem-synth; if the raw pitch accuracy falls below PESTO's 96.1% and 95.5% respectively, as the paper's own Table 1 suggests, the 'original SWIPE surpasses all self-supervised pitch detection' claim does not hold for the ERB configuration.

Watch

Extended reading notes

Core claim

The central claim is that the SWIPE algorithm—which scores each pitch candidate by the normalized inner product of a cosine-lobed harmonic kernel with the square-root magnitude spectrum of the input—has been consistently underreported as a baseline, and that in its standard SWIPE' form with mel-scale spectral sampling it is more accurate than the state-of-the-art self-supervised network PESTO on the MIR-1K and MDB-stem-synth datasets. In the authors' measurements, SWIPE reaches 96.2% and 96.1% raw pitch accuracy on those two sets, compared with PESTO's 96.1% and 95.5%, respectively, when each model is trained on the same dataset. The paper further claims that these same SWIPE scores are a sufficient input representation for self-supervised learning: a Toeplitz-only encoder with 647 parameters, essentially learning a reweighting of the SWIPE scores, matches or exceeds PESTO's accuracy on same-dataset evaluation and is more robust to white noise. In supervised training, a SWIPE-fed network generalizes better to an unseen singing dataset than a raw-audio network with 6.6 million parameters.

Load-bearing premise

The claim that original SWIPE beats PESTO relies on sampling the spectrum on the mel scale; with the ERB scale (the original SWIPE recommendation for speech) the same implementation scores below PESTO on both datasets.

Editorial extensions

If this is right

SWIPE should replace the weak baseline numbers that previous self-supervised papers reported; its true accuracy on music benchmarks is near or above 96% raw pitch accuracy.
Self-supervised pitch networks can be reduced from 28,900 parameters to 647 parameters by using SWIPE scores as input, without a loss of same-dataset accuracy and with better noise robustness.
The window size of the SWIPE frontend can be reduced at inference time without retraining, trading accuracy for latency down to tens of milliseconds, which suits real-time streaming use.
Supervised pitch estimators using SWIPE features show better cross-domain generalization to unseen singing voices than raw-audio networks, despite having far fewer parameters.
Comparisons of neural pitch estimators against DSP baselines should report the baseline implementation's search range and frequency scale, since those choices can swing accuracy by more than 20 percentage points.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper's single-pitch benchmarks, a natural extension is to test SWIPE-frontended networks in multi-pitch or polyphonic settings, since the kernels score one candidate at a time and the Toeplitz layer assumes one pitch per frame.
The paper's finding that feature-space augmentation of SWIPE scores degrades accuracy, while time-domain augmentation helps, hints that hand-crafted DSP features are sensitive to naive perturbations and that augmentation design must respect the feature's generative structure.
If SWIPE scores remain this informative on larger, noisier corpora, they could serve as cheap pseudo-label teachers for training or distilling neural pitch estimators, reducing the need for labeled data even further.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The genuinely new thing here is a clean re-evaluation of SWIPE using a careful implementation, plus the demonstration that SWIPE scores work as a neural frontend for both supervised and self-supervised pitch estimators. On the self-supervised side, a 647-parameter Toeplitz-only model (SWIPE-tiny) matches or beats PESTO on same-dataset RPA and does far better on cross-dataset in the MDB-to-MIR-1K direction. That holds up on reading. The paper is unusually transparent about implementation choices—their Table 1 reports both ERB and mel sampling, and they release code and trained models.

The stress-test note is partially right. The 'SWIPE in its original form surpasses PESTO' claim depends on the mel-sampled variant; with ERB sampling, the authors' own numbers on MIR-1K (95.7%) and MDB (94.0%) drop below PESTO's same-dataset numbers (96.1% and 95.5%). That said, the paper openly reports this and motivates mel as the original SWIPE recommendation for musical instruments. MIR-1K is singing voice, which sits between speech and music in that framework. I would not call the claim circular or fabricated—just imprecise. It should be re-scoped to 'a mel-sampled SWIPE implementation,' and the margins (0.1 to 0.6 percentage points) need confidence intervals or multiple seeds before 'surpass' is the right verb. The absence of any error bars or repeated training runs across the whole paper is a moderate concern, especially where a 1 pp gap separates 'competitive' from 'better.'

The broader 'SWIPE has been underreported' message is robust: even the ERB version easily beats the previously reported baseline values (86.6/90.7) and PYIN. So the core claim—that SWIPE deserves a better place in the evaluation stack—does not depend on the fragile mel-vs-ERB distinction.

This paper is for audio/MIR researchers building pitch estimators, anyone benchmarking self-supervised methods, and people who want a reminder that DSP baselines are not dead. It is a well-written, honest empirical paper. I would send it to peer review, with the request that the abstract and conclusion be rephrased to 'mel-scaled SWIPE' and that uncertainty be reported explicitly.

Referee Report

3 major / 6 minor

Summary. The paper investigates using SWIPE scores as an audio frontend for neural pitch estimators, replacing the CQT representation used in the PESTO architecture. The authors evaluate supervised (CQT-sup, SWIPE-sup) and self-supervised (CQT-tiny, SWIPE-full, SWIPE-tiny) models on MDB-stem-synth, MIR-1K, and PTDB-TUG. They report that SWIPE-based frontends reduce network size by an order of magnitude without performance loss, that a 647-parameter SWIPE-tiny model matches or exceeds PESTO accuracy with better noise robustness, and that the SWIPE algorithm in its original form outperforms the PESTO self-supervised model on the evaluated datasets.

Significance. If the claims are upheld, the paper delivers a notable result: a 2008 hand-crafted DSP algorithm surpassing a trained 28.9k-parameter self-supervised network on monophonic pitch accuracy, and a parameter-efficient SWIPE-tiny model with superior noise robustness. The paper is transparent about implementation details, releases code, and includes direct comparisons with strong baselines. However, the small accuracy margins and the sensitivity of the central comparison to the choice of frequency scale (mel vs. ERB) mean the headline claims should be interpreted with caution.

major comments (3)

[Abstract; §3.1; §5.2] The claim that 'SWIPE in its original form surpasses the accuracy of the state of the art in self-supervised neural pitch detection (PESTO)' is not supported by the reported configuration. Section 3.1 describes an implementation that uses mel-scale sampling, SWIPE' kernels, and a candidate set of 295 bins; Section 2.1.1 notes that the original SWIPE paper recommends ERB sampling for speech and mel sampling for musical instruments. The datasets MIR-1K and MDB-stem-synth contain vocals, so the original recommendation would be ERB for the vocal portions. Table 1 shows that the authors' implementation with ERB sampling achieves 95.7% RPA on MIR-1K, below PESTO's 96.1%. The superiority claim therefore rests on a hand-selected mel scale rather than on the original algorithm's defaults, and the 'original form' wording should be revised or justified.
[Table 3; §5.2] The reported margins over PESTO are very small: 96.2% vs. 96.1% on MIR-1K and 96.1% vs. 95.5% on MDB for the SWIPE baseline, and similar margins for SWIPE-tiny. No confidence intervals, standard deviations, or multiple training seeds are reported for any of the neural models. Given the stochastic nature of self-supervised training, the conclusion that SWIPE 'outperforms' PESTO is not statistically substantiated. The authors should provide error bars or a significance test, or temper the wording to 'marginally exceeds' in these comparisons.
[§5.2, first paragraph] The statement that 'the original SWIPE algorithm outperforms all work on self-supervised monophonic pitch detection published to date' goes beyond the evidence, which is a comparison with PESTO (and indirect references to SPICE via Table 1). A claim about all published work would require a broader set of baselines or a survey. The claim should be restricted to the methods compared in this paper.

minor comments (6)

[§5.3] The text states that for fmin=27.5 Hz the next-longest power-of-two window corresponds to 327 ms, but the largest power-of-two window in Table 4 is 16384 samples, which at 44.1 kHz is 371.5 ms; the 327 ms value appears to be a typo.
[§3.1] The use of the Slaney-style mel scale is mentioned without a citation; please add the appropriate reference (e.g., Slaney, 1998).
[Table 1 and text] The dataset is referred to as 'MDB' in Table 1 but 'MDB-stem-synth' in the text; please use consistent nomenclature.
[§2.3.1, Eq. (2)] The definition of φ is written as mapping y to (α, α^2, ..., α^{|C|}) y; the notation is confusing because the expression is a scalar while the codomain is R. Clarify that this is a dot product.
[Author block] The corresponding author email (d.sudholt@qmul.ac.uk) does not match the listed authors (David Marttila, Joshua D. Reiss); please correct.
[§4.2] The SWIPE baseline is described as not using peak refinement, but Section 3.1 notes that SPTK uses parabolic interpolation; clarify whether the authors' SWIPE implementation applies any interpolation when computing scores.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical comparisons against external baselines, with no fitted parameter renamed as a prediction and no load-bearing self-citation.

full rationale

The paper is an empirical study comparing a hand-crafted DSP feature (SWIPE) and learned frontends on standard pitch-estimation benchmarks. The central claims—that SWIPE scores improve neural pitch estimators and that a carefully implemented SWIPE can outperform PESTO—are evaluated through held-out RPA, F-Score, and OA measurements on MIR-1K, MDB-stem-synth, and PTDB-TUG. No equation in the paper defines a key result in terms of the quantity it is claimed to predict, and no parameter is fit to a subset and then reported as a prediction of that same subset. The SWIPE implementation's sampling scale (mel vs. ERB) is a design choice that affects results, but it is not an input that is later recovered as an output; Table 1 explicitly reports both variants and the comparison to SPTK, making the dependence transparent. The self-supervised PESTO baseline is cited from prior external work, not from the authors' own prior results. References to the original SWIPE paper and SPTK are standard citations to external code and publications. The only arguable weakness—that the 'original form' headline depends on choosing the mel scale rather than the ERB scale recommended for speech—is a matter of result robustness and interpretation, not circularity, because the comparison would still be an empirical claim even if the choice were contested. Therefore no circular step can be exhibited, and the analysis yields a score of 0.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central empirical results rest on implementation choices (mel scale, SWIPE' variant, score threshold) and standard benchmarking assumptions. No genuinely new entity is introduced; the free parameters are hand choices rather than fitted values, but the mel-scale choice is load-bearing for the headline comparison.

free parameters (4)

Frequency sampling scale = mel (Slaney), not ERB
The authors sample SWIPE spectra on a mel scale for all datasets; switching to ERB drops MIR-1K RPA from 96.2% to 95.7% and MDB from 96.1% to 94.0%, below PESTO. This hand choice is load-bearing for the headline claim.
SWIPE kernel variant = SWIPE' (non-prime harmonics removed)
They use the prime-harmonic variant and call it 'original form'; the classic SWIPE includes all integer harmonics. The contribution of this variant to the high scores is not isolated.
Score threshold = 0
The threshold for pitch candidates to be considered is set to 0; different thresholds would change voiced/unvoiced behavior.
Number of spectrum samples = 1024
SWIPE scores are computed at 1024 mel-spaced frequencies; the dependence of results on this resolution is not explored.

assumptions (3)

standard math SWIPE score formula (Eq. 1) correctly measures pitch similarity
The paper relies on the SWIPE score as defined by Camacho and Harris; this is prior work, not re-derived.
domain assumption Datasets MIR-1K, MDB-stem-synth, PTDB-TUG are representative benchmarks for monophonic pitch estimation
The claim of outperforming self-supervised SOTA is based on these three datasets; the paper does not test on noisy speech or multi-pitch material.
domain assumption PESTO baseline was trained faithfully
The comparison assumes the PESTO numbers reflect proper training; the authors report they replicate the training setup, but no code or checkpoints for the baseline are provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Improving Neural Pitch Estimation with SWIPE Kernels." pith.science (2026). https://pith.science/paper/DKFZ425T

@misc{pith2026250711233,
  author       = {Pith},
  title        = {Pith review of: Improving Neural Pitch Estimation with SWIPE Kernels},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DKFZ425T}},
  note         = {Machine review of arXiv:2507.11233}
}

read the original abstract

Neural networks have become the dominant technique for accurate pitch and periodicity estimation. Although a lot of research has gone into improving network architectures and training paradigms, most approaches operate directly on the raw audio waveform or on general-purpose time-frequency representations. We investigate the use of Sawtooth-Inspired Pitch Estimation (SWIPE) kernels as an audio frontend and find that these hand-crafted, task-specific features can make neural pitch estimators more accurate, robust to noise, and more parameter-efficient. We evaluate supervised and self-supervised state-of-the-art architectures on common datasets and show that the SWIPE audio frontend allows for reducing the network size by an order of magnitude without performance degradation. Additionally, we show that the SWIPE algorithm on its own is much more accurate than commonly reported, outperforming state-of-the-art self-supervised neural pitch estimators.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 32 canonical work pages

[1]

Improving Neural Pitch Estimation with SWIPE Kernels

INTRODUCTION Pitch plays a central role in how humans perceive sound. Consequently, pitch estimation is a fundamental task in many music, speech and audio processing pipelines. While pitch is a psychoacoustic phenomenon, it closely corre- lates to the signal processing concept of the fundamental frequency f0. Recent literature commonly uses the term “pitc...

work page Pith review arXiv 2025
[2]

BACKGROUND 2.1 SWIPE The Sawtooth Waveform Inspired Pitch Estimator (SWIPE) [3] estimates pitch by identifying the fundamen- tal frequency of a sawtooth waveform whose spectrum best matches that of the input signal. To achieve this, it constructs spectral kernels for a number of discrete pitch candidates, and assigns a score to each pitch candidate by mea...

work page
[3]

METHODS The core insight of this work is that SWIPE scores encode rich pitch information and are thus well suited as an audio frontend for neural pitch estimation in both supervised and self-supervised settings. This section first covers our im- plementation of SWIPE in detail, and then describes how we adapt supervised and self-supervised neural pitch es...

work page
[4]

The au- dio is re-synthesized from its f0 annotations, which means that the f0 annotations are perfect

EXPERIMENTAL SETUP 4.1 Datasets Our experiments use three f0-annotated datasets that are commonly used for training and benchmarking pitch de- tectors: MDB-stem-synth [21] contains 230 solo tracks (418 minutes total) of instrument sounds and vocals. The au- dio is re-synthesized from its f0 annotations, which means that the f0 annotations are perfect. It ...

work page
[5]

tiny” refers to the Toeplitz-only encoder and “full

RESULTS AND DISCUSSION We report separate experimental results for the supervised and self-supervised approaches, in each case closely repli- cating the training setup of the baselines (FCNF0++ and PESTO, respectively) to assess the impact of using SWIPE scores as an audio frontend. 5.1 Supervised Models We refer to the two proposed supervised models (see...

work page
[6]

CONCLUSION We investigated the potential of combining the SWIPE al- gorithm with neural pitch estimation. We adapted estab- lished supervised and self-supervised training techniques to use SWIPE scores as an audio frontend and obtained accurate, efficient, robust and flexible pitch estimators. We demonstrated that the potential of SWIPE has been significa...

work page
[7]

The authors would like to thank the anonymous reviewers for their valuable feedback which significantly improved this paper

ACKNOWLEDGMENTS This work was supported by UK Research and Innovation [grant number EP/S022694/1]. The authors would like to thank the anonymous reviewers for their valuable feedback which significantly improved this paper

work page
[8]

Cepstrum Pitch Determination,

A. M. Noll, “Cepstrum Pitch Determination,” The Journal of the Acoustical Society of America , vol. 41, no. 2, pp. 293–309, Feb. 1967

work page 1967

Show all 33 references

[9]

Fundamental fre- quency estimation of musical signals using a two-way mismatch procedure,

R. C. Maher and J. W. Beauchamp, “Fundamental fre- quency estimation of musical signals using a two-way mismatch procedure,” The Journal of the Acoustical Society of America , vol. 95, no. 4, pp. 2254–2263, 1994

1994
[10]

A sawtooth waveform inspired pitch estimator for speech and music,

A. Camacho and J. G. Harris, “A sawtooth waveform inspired pitch estimator for speech and music,” The Journal of the Acoustical Society of America , vol. 124, no. 3, pp. 1638–1652, Sep. 2008

2008
[11]

PEFAC - A Pitch Es- timation Algorithm Robust to High Levels of Noise,

S. Gonzalez and M. Brookes, “PEFAC - A Pitch Es- timation Algorithm Robust to High Levels of Noise,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing , vol. 22, no. 2, pp. 518–530, Feb. 2014

2014
[12]

YIN, a Funda- mental Frequency Estimator for Speech and Music,

A. de Cheveigné and H. Kawahara, “YIN, a Funda- mental Frequency Estimator for Speech and Music,” The Journal of the Acoustical Society of America , vol. 111, no. 4, pp. 1917–1930, 2002

1917
[13]

A smarter way to find pitch,

P. McLeod and G. Wyvill, “A smarter way to find pitch,” in International Computer Music Conference (ICMC), 2005

2005
[14]

CREPE: A Convolutional Representation for Pitch Estimation,

J. W. Kim, J. Salamon, P. Li, and J. P. Bello, “CREPE: A Convolutional Representation for Pitch Estimation,” in Proceedings of the IEEE International Confer- ence on Acoustics, Speech, and Signal Processing (ICASSP), 2018

2018
[15]

Fully-Convolutional Net- work for Pitch Estimation of Speech Signals,

L. Ardaillon and A. Roebel, “Fully-Convolutional Net- work for Pitch Estimation of Speech Signals,” inInter- speech 2019, 2019

2019
[16]

Cross-domain Neural Pitch and Periodicity Estima- tion,

M. Morrison, C. Hsieh, N. Pruyne, and B. Pardo, “Cross-domain Neural Pitch and Periodicity Estima- tion,” http://arxiv.org/abs/2301.12258, Jun. 2023

2023 arXiv
[17]

YOLOPitch: A Time-Frequency Dual-Branch YOLO Model for Pitch Estimation,

X. Li, H. Huang, Y . Hu, L. He, J. Zhang, and Y . Wang, “YOLOPitch: A Time-Frequency Dual-Branch YOLO Model for Pitch Estimation,” in Interspeech 2024 , 2024

2024
[18]

Pitch estima- tion via self-supervision,

B. Gfeller, C. Frank, D. Roblek, M. Sharifi, M. Tagliasacchi, and M. Velimirovi ´c, “Pitch estima- tion via self-supervision,” in 2020 IEEE International Conference on Acoustics, Speech and Signal Process- ing (ICASSP), 2020

2020
[19]

SPICE: Self- Supervised Pitch Estimation,

B. Gfeller, C. Frank, D. Roblek, M. Sharifi, M. Tagliasacchi, and M. Velimirovic, “SPICE: Self- Supervised Pitch Estimation,” IEEE/ACM Transac- tions on Audio Speech and Language Processing , vol. 28, pp. 1118–1128, 2020

2020
[20]

Self-Supervised Pitch Detection by In- verse Audio Synthesis,

J. Engel, R. Swavely, A. Roberts, L. H. Hantrakul, and C. Hawthorne, “Self-Supervised Pitch Detection by In- verse Audio Synthesis,” Workshop on Self-Supervision in Audio and Speech at the 37th International Confer- ence on Machine Learning (ICML 2020) , 2020

2020
[21]

PESTO: Pitch estimation with self-supervised transposition-equivariant objective,

A. Riou, S. Lattner, G. Hadjeres, and G. Peeters, “PESTO: Pitch estimation with self-supervised transposition-equivariant objective,” in Proceedings of the 24th International Society for Music Information Retrieval Conference (ISMIR), 2023

2023
[22]

hf0: A Hybrid Pitch Extraction Method for Multimodal V oice,

P. Rengaswamy, M. G. Reddy, K. S. Rao, and P. Das- gupta, “hf0: A Hybrid Pitch Extraction Method for Multimodal V oice,”Circuits, Systems, and Signal Pro- cessing, vol. 40, no. 1, pp. 262–275, Jan. 2021

2021
[23]

HAEPF: Hybrid approach for estimating pitch frequency in the presence of reverberation,

E. S. Hassan, B. Neyazi, H. S. Seddeq, A. Z. Mah- moud, A. S. Oshaba, A. El-Emary, and F. E. Abd El- Samie, “HAEPF: Hybrid approach for estimating pitch frequency in the presence of reverberation,” Multime- dia Tools and Applications, vol. 83, no. 32, pp. 77 489– 77 508, Feb. 2024

2024
[24]

Pytorch: An imper- ative style, high-performance deep learning library,

A. Paszke, S. Gross, F. Massa, A. Lerer, J. Brad- bury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, “Pytorch: An imper- ative style, hig...

2019
[25]

Derivation of au- ditory filter shapes from notched-noise data,

B. R. Glasberg and B. C. Moore, “Derivation of au- ditory filter shapes from notched-noise data,” Hearing Research, vol. 47, no. 1, pp. 103–138, 1990

1990
[26]

SPTK4: An open-source software toolkit for speech signal processing,

T. Yoshimura, T. Fujimoto, K. Oura, and K. Tokuda, “SPTK4: An open-source software toolkit for speech signal processing,” in 12th ISCA Speech Synthesis Workshop (SSW 2023), 2023, pp. 211–217

2023
[27]

librosa: Audio and music signal analysis in python

B. McFee, C. Raffel, D. Liang, D. P. Ellis, M. McVicar, E. Battenberg, and O. Nieto, “librosa: Audio and music signal analysis in python.” in Proceedings of the 14th Python in Science Conference, 2015

2015
[28]

An analysis/synthesis framework for automatic f0 annotation of multitrack datasets

J. Salamon, R. M. Bittner, J. Bonada, J. J. Bosch, E. Gómez, and J. P. Bello, “An analysis/synthesis framework for automatic f0 annotation of multitrack datasets.” in Proceedings of the 18th International So- ciety for Music Information Retrieval Conference (IS- MIR), 2017

2017
[29]

A pitch tracking corpus with evaluation on multip- itch tracking scenario

G. Pirker, M. Wohlmayr, S. Petrik, and F. Pernkopf, “A pitch tracking corpus with evaluation on multip- itch tracking scenario.” inInterspeech, 2011, pp. 1509– 1512

2011
[30]

On the improvement of singing voice separation for monaural recordings us- ing the mir-1k dataset,

C.-L. Hsu and J.-S. R. Jang, “On the improvement of singing voice separation for monaural recordings us- ing the mir-1k dataset,” IEEE Transactions on Audio, Speech, and Language Processing , vol. 18, no. 2, pp. 310–319, 2010

2010
[31]

pYIN: A fundamental fre- quency estimator using probabilistic threshold distri- butions,

M. Mauch and S. Dixon, “pYIN: A fundamental fre- quency estimator using probabilistic threshold distri- butions,” in 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2014, pp. 659–663

2014
[32]

Mir_eval: A transparent implementation of common mir metrics

C. Raffel, B. McFee, E. J. Humphrey, J. Salamon, O. Nieto, D. Liang, D. P. Ellis, and C. C. Raffel, “Mir_eval: A transparent implementation of common mir metrics.” in Proceedings of the 15th International Society for Music Information Retrieval Conference (ISMIR), 2014

2014
[33]

Adam: A method for stochastic optimization,

D. P. Kingma and J. L. Ba, “Adam: A method for stochastic optimization,” in 3rd International Confer- ence on Learning Representations (ICLR) , 2015

2015

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Improving Neural Pitch Estimation with SWIPE Kernels

INTRODUCTION Pitch plays a central role in how humans perceive sound. Consequently, pitch estimation is a fundamental task in many music, speech and audio processing pipelines. While pitch is a psychoacoustic phenomenon, it closely corre- lates to the signal processing concept of the fundamental frequency f0. Recent literature commonly uses the term “pitc...

work page Pith review arXiv 2025

[2] [2]

BACKGROUND 2.1 SWIPE The Sawtooth Waveform Inspired Pitch Estimator (SWIPE) [3] estimates pitch by identifying the fundamen- tal frequency of a sawtooth waveform whose spectrum best matches that of the input signal. To achieve this, it constructs spectral kernels for a number of discrete pitch candidates, and assigns a score to each pitch candidate by mea...

work page

[3] [3]

METHODS The core insight of this work is that SWIPE scores encode rich pitch information and are thus well suited as an audio frontend for neural pitch estimation in both supervised and self-supervised settings. This section first covers our im- plementation of SWIPE in detail, and then describes how we adapt supervised and self-supervised neural pitch es...

work page

[4] [4]

The au- dio is re-synthesized from its f0 annotations, which means that the f0 annotations are perfect

EXPERIMENTAL SETUP 4.1 Datasets Our experiments use three f0-annotated datasets that are commonly used for training and benchmarking pitch de- tectors: MDB-stem-synth [21] contains 230 solo tracks (418 minutes total) of instrument sounds and vocals. The au- dio is re-synthesized from its f0 annotations, which means that the f0 annotations are perfect. It ...

work page

[5] [5]

tiny” refers to the Toeplitz-only encoder and “full

RESULTS AND DISCUSSION We report separate experimental results for the supervised and self-supervised approaches, in each case closely repli- cating the training setup of the baselines (FCNF0++ and PESTO, respectively) to assess the impact of using SWIPE scores as an audio frontend. 5.1 Supervised Models We refer to the two proposed supervised models (see...

work page

[6] [6]

CONCLUSION We investigated the potential of combining the SWIPE al- gorithm with neural pitch estimation. We adapted estab- lished supervised and self-supervised training techniques to use SWIPE scores as an audio frontend and obtained accurate, efficient, robust and flexible pitch estimators. We demonstrated that the potential of SWIPE has been significa...

work page

[7] [7]

The authors would like to thank the anonymous reviewers for their valuable feedback which significantly improved this paper

ACKNOWLEDGMENTS This work was supported by UK Research and Innovation [grant number EP/S022694/1]. The authors would like to thank the anonymous reviewers for their valuable feedback which significantly improved this paper

work page

[8] [8]

Cepstrum Pitch Determination,

A. M. Noll, “Cepstrum Pitch Determination,” The Journal of the Acoustical Society of America , vol. 41, no. 2, pp. 293–309, Feb. 1967

work page 1967