Pith. sign in

REVIEW 2 major objections 4 minor 39 references

SLASH: Self-Supervised Speech Pitch Estimation Leveraging DSP-derived Absolute Pitch

T0 review · 2 major / 4 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 SLASH shows that adding DSP-computed absolute pitch to self-supervised training lets a pitch estimator beat DSP and SSL baselines and approach supervised accuracy on MIR-1K, all without ground-truth F0 labels.

desk verdict A clever self-supervised pitch method with a real novelty in the pseudo-spectrogram objective, but the headline claim against SSL baselines rests on an uncontrolled PESTO comparison that needs fixing. read the letter →

arxiv 2507.17208 v1 pith:IOVWN7ES submitted 2025-07-23 eess.AS

Ryo Terashima , Yuma Shirahata , Masaya Kawamura This is my paper
classification eess.AS
keywords pitchestimationself-supervisedlearningfundamentalfrequencydifferentiableDSPsubharmonicsummationvoiced/unvoiceddetectionaperiodicity
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

SLASH asks whether a self-supervised pitch estimator can learn absolute pitch values without any ground-truth F0 labels. The paper's answer is yes: two DSP-based objectives, a subharmonic-summation pitch guide and a differentiable spectrogram-synthesis loss, give the model an absolute reference while pitch-shift consistency provides relative structure. Trained on unlabeled speech, SLASH beats the DSP baselines DIO and Harvest and the SSL baseline PESTO on pitch accuracy, log-F0 RMSE, and V/UV error, and reaches accuracy comparable to the supervised CREPE. The practical upshot is that high-quality pitch tracking can be built from raw audio and classical signal-processing priors alone, with no manual annotation.

What carries the argument

The load-bearing object is a differentiable pseudo-spectrogram generator that turns the predicted F0 $p_t$ into a periodic-excitation spectrum $E^*_p$ using a triangle-wave phase model $\Phi_{t,k}=\frac{f_s}{2 p_t K} k$, with additive small noise to avoid zeros. The generator then forms the full pseudo spectrogram $S^*$ by combining this excitation with the spectral envelope and aperiodicity, and the loss compares the fine-structure spectra $\psi(S^*)$ and $\psi(S)$ under an $\ell^1$ norm masked by voicing. Because the spectrogram is computed directly from F0, waveform synthesis—the known source of instability in differentiable-DSP gradient descent—is skipped, which the paper argues makes F0 optimization viable. A pitch guide $G$, computed by subharmonic summation on the fine-structure spectrum and used in a hinge loss, supplies a coarse absolute reference, while a differentiable WORLD-style synthesizer with generalized energy distance refines aperiodicity and voicing.

What would settle it

Retrain PESTO from scratch on exactly the 700 MIR-1K phrases used to train SLASH (or any split that excludes the 250 test phrases) and re-run Table 1; if its accuracy rises to or above SLASH's, the claimed edge over the SSL baseline is an artifact of training-set overlap.

Watch

Extended reading notes

Core claim

The central claim is that self-supervised pitch training does not have to settle for pitch-shift equivariance as its only learning signal. SLASH obtains a frame-level prior over F0 using subharmonic summation on the fine-structure spectrum, and a second loss that reconstructs a periodic-excitation spectrogram directly from the predicted F0, so gradient descent can align harmonics with the target. Because the pseudo-spectrogram path bypasses waveform generation, the authors argue, F0 optimization by gradient descent becomes stable enough to sharpen predictions beyond what the pitch-guide loss alone achieves. The experiments on MIR-1K support this: SLASH reaches 0.969 raw pitch accuracy (50c) with a log-F0 RMSE of 0.018 when trained on LibriTTS-R and 0.967/0.017 when trained on MIR-1K, surpassing PESTO (0.962/0.057) and DIO (0.943/0.030), and staying competitive with the supervised CREPE while using no manual labels.

Load-bearing premise

The head-to-head comparison against the SSL baseline assumes that the public PESTO model was not trained on the same 250 MIR-1K phrases used for testing, since the paper states only that PESTO was trained on MIR-1K and does not report whether those phrases were held out.

Editorial extensions

If this is right

  • Absolute pitch can be learned self-supervised: the ablation shows removing the pitch guide and pseudo-spectrogram losses collapses RPA to near zero, while adding the guide alone only partially recovers accuracy.
  • The method's aperiodicity and V/UV predictions, optimized via generalized energy distance, make the same model useful for vocoder analysis, not just pitch tracking.
  • Out-of-domain generalization is strong: training on LibriTTS-R read speech yields pitch and V/UV accuracy on MIR-1K singing voice nearly equal to training on MIR-1K itself.
  • Noise robustness comes from the waveform augmentation and associated losses: at 0 dB SNR the model degrades minimally while DSP and other baselines deteriorate substantially.

Reading between the lines

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

  • The same pseudo-spectrogram trick—bypassing waveform generation to stabilize gradient descent on a signal parameter—should transfer to other differentiable synthesis settings, such as estimating spectral envelopes or aperiodic components directly from a loss on a target spectrogram, not just F0.
  • Because SLASH trained on LibriTTS-R (read speech) already performs near the MIR-1K-trained version on singing voice, large unlabeled speech corpora may be enough to build pitch estimators for musical vocals, a regime where labeled F0 is scarce.
  • The paper's absolute-pitch losses could be combined with other pretext tasks, such as speaker or emotion prediction, potentially giving SSL speech models a richer notion of prosody without any labeled pitch.
  • A clean evaluation in which every baseline is retrained on identical, non-overlapping data splits would determine how much of the reported margin over PESTO is a genuine architectural advantage rather than a dataset artifact.
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.

Referee Report

2 major / 4 minor

Summary. SLASH is a self-supervised pitch estimation method for speech that combines relative-pitch consistency (pitch-shift equivariance) with two DSP-derived absolute-pitch cues: a subharmonic-summation-based pitch guide used in a hinge loss, and a gradient-descent objective that compares a directly synthesized periodic-excitation spectrogram with the target log fine-structure spectrum. The model also predicts band aperiodicity and V/UV flags through a differentiable WORLD-like synthesizer and a generalized energy distance, and it is trained with noise/volume augmentation. Evaluated on 250 MIR-1K phrases against DIO, Harvest, PESTO, and CREPE, the paper reports that SLASH outperforms the DSP and SSL baselines on RPA, log-F0 RMSE, and V/UV error, while being competitive with the supervised CREPE; ablations show that removing the absolute-pitch losses severely degrades pitch accuracy.

Significance. If the reported results hold, the paper makes a useful contribution by showing that absolute pitch information can be injected into a self-supervised pitch estimator without ground-truth labels, using differentiable DSP and a novel spectrogram-generation shortcut that avoids backpropagating through waveform synthesis. The ablation study (Table 2) is informative and supports the claim that both the pitch-guide loss and the pseudo-spectrogram loss contribute. The method is internally consistent: all training losses are computed from the input itself, with no ground-truth labels or fitted constants encoding the target pitch. The main risk is empirical: the head-to-head comparison with the SSL baseline PESTO may not be a clean held-out comparison, and no uncertainty estimates are provided for the reported metrics.

major comments (2)
  1. [Sec. 3.1 (Comparison models) and Table 1] The only SSL baseline, PESTO, is the official pretrained model reported to be trained on the MIR-1K dataset, while the test set is 250 randomly selected MIR-1K phrases. The paper does not state whether those 250 phrases were excluded from PESTO's training data. If they were not, the Table 1 margins (RPA 0.967 vs 0.962, log-F0 RMSE 0.017 vs 0.057) are not a valid held-out comparison, and the abstract's claim that SLASH outperforms SSL-based pitch estimation methods is unsupported. Please retrain PESTO on the same 700-phrase training split used for SLASH(MIR-1K), or evaluate on an independent corpus, and report the overlap explicitly.
  2. [Table 1 and Fig. 4] The reported numbers come from a single experimental run or a single test split, with no confidence intervals, standard deviations, or significance tests. Several decisive margins are small (e.g., RPA 0.969 vs 0.962, V/UV ER 0.033 vs 0.052), and the noise-robustness figure compares methods without error bars. At minimum, report multiple runs with different random seeds or bootstrap confidence intervals over the 250 test phrases, and a significance test for the main comparisons.
minor comments (4)
  1. [Sec. 3.1 (Datasets)] Please state explicitly that the 250 test phrases are disjoint from the 700 training and 50 validation phrases used for SLASH(MIR-1K), and report any song- or singer-level overlap between the training and test portions of MIR-1K.
  2. [Sec. 2.4, Eq. (4)] The exponent in 'max(X, ε)2' is ambiguous; write max(X, ε)^2. Also, the symbols E*_p and E*_t in Figure 3 and the surrounding text should be used consistently.
  3. [Sec. 2.2 and Sec. 2.3] There are two typos: 'an b-dimensional Band Aperiodicity' should be 'a b-dimensional Band Aperiodicity', and 'course F0 structure' should be 'coarse F0 structure'.
  4. [Sec. 3.2.2 and Fig. 4] The SNR conditions in the noise-robustness experiment are not fully defined; specify whether SNR is computed utterance-level, how noise is scaled, and whether the same noise realizations are used across models, so that the experiment is replicable.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: SLASH's losses are internal-consistency objectives on the input waveform; the DSP pitch guide is an openly disclosed training prior and is not fed to the model at inference.

full rationale

SLASH's training objectives are all self-supervised and input-derived. Lcons (Eq. 1) compares pitch estimates of original and pitch-shifted CQTs, learning relative pitch; Lg (Eq. 3) aligns the predicted F0 distribution with a DSP-derived SHS prior G, which the paper transparently describes as a 'pitch guide'; Lpseudo (Eq. 6) and Lrecon (Eq. 8) are analysis-by-synthesis losses that compare pseudo/target fine-structure spectra and reconstructed spectrograms to the same input spectrogram. None of these losses is defined in terms of the ground-truth pitch labels, and none is a fitted parameter later reported as a prediction. The pitch guide is not an input to the Pitch Encoder at inference: during inference only the CQT is fed to the network, so the model cannot simply copy G. The paper does not rely on self-citations or imported uniqueness theorems to justify its architecture; the only close architectural borrowing is from NANSY++ [12], which is cited as prior work and is not load-bearing for the self-supervised claim. The ablation in Table 2 shows that removing the absolute-pitch losses collapses RPA to 0.000, confirming that those losses carry the absolute-pitch information rather than being redundant with the relative-pitch objective. The comparison against PESTO on MIR-1K has a potential benchmark-validity limitation: Sec. 3.1 states that the test set is '250 randomly selected phrases from the MIR-1K dataset' and that the official PESTO model 'is trained on the MIR-1K dataset,' without reporting whether those 250 phrases were excluded from PESTO's training data. That is a correctness/fairness concern about a third-party checkpoint's training split, not a circularity of SLASH's derivation. Overall, no step in the paper reduces by definition or by self-citation to its own inputs.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim depends on a handful of hand-chosen hyperparameters and standard speech-modeling domain assumptions. No data-fitted constants encode the target pitch, so the circularity burden is low.

free parameters (5)
  • hinge margin m in pitch guide loss (eq. 3) = 0.5
    Sets how much the pitch prediction must overlap the DSP guide before the loss is zero; hand-chosen, no sensitivity analysis reported.
  • GED weight alpha (eq. 8) = 0.1
    Controls the repulsive term in the reconstruction loss; hand-chosen.
  • V/UV threshold theta (eq. 9) = 0.5
    Binarizes the voicing decision; hand-chosen.
  • epsilon in pseudo excitation (eq. 4) = 0.001
    Noise floor added to the excitation spectrum; hand-chosen.
  • loss weights (Lcons, Lpseudo, Lrecon, Lg, Lg-shift, Laug, Lg-aug, Lap) = 10, 10, 5, 1, 1, 1, 1, 1
    Relative weighting of all objective terms; hand-chosen, no tuning procedure reported.
assumptions (5)
  • domain assumption Speech can be represented by the source-filter model: waveform decomposes into F0, spectral envelope, and aperiodicity.
    Used throughout Sec. 2.1 and 2.5 to justify the analysis-synthesis architecture.
  • domain assumption The fine structure spectrum psi(S) after lag-window envelope removal contains harmonic peaks at multiples of F0.
    The pseudo-spectrogram loss (eq. 6) aligns the generated excitation peaks with psi(S); this is standard for harmonic speech, but not true for creaky or very noisy segments.
  • domain assumption SHS-based pitch guide G is a reliable enough prior for absolute pitch that the hinge loss (eq. 3) can steer learning.
    If SHS has systematic errors (e.g., octave errors), the guide could bias the model; the hinge margin is meant to tolerate disagreement.
  • domain assumption Shifting the CQT bins approximates a pitch shift of the underlying audio.
    Used for the pitch consistency loss in Sec. 2.2; standard in PESTO/SPICE, but ignores spectral envelope changes under true pitch shift.
  • ad hoc to paper The triangle-wave model of the periodic excitation is a sufficient approximation for gradient-based F0 learning.
    Eq. (4) defines a hand-crafted spectral excitation whose peak spacing is controlled by p_t; whether its harmonic amplitude shape matches real speech is not validated, though log-domain loss and empirical results mitigate this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SLASH: Self-Supervised Speech Pitch Estimation Leveraging DSP-derived Absolute Pitch." pith.science (2026). https://pith.science/paper/IOVWN7ES

@misc{pith2026250717208,
  author       = {Pith},
  title        = {Pith review of: SLASH: Self-Supervised Speech Pitch Estimation Leveraging DSP-derived Absolute Pitch},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IOVWN7ES}},
  note         = {Machine review of arXiv:2507.17208}
}
read the original abstract

We present SLASH, a pitch estimation method of speech signals based on self-supervised learning (SSL). To enhance the performance of conventional SSL-based approaches that primarily depend on the relative pitch difference derived from pitch shifting, our method incorporates absolute pitch values by 1) introducing a prior pitch distribution derived from digital signal processing (DSP), and 2) optimizing absolute pitch through gradient descent with a loss between the target and differentiable DSP-derived spectrograms. To stabilize the optimization, a novel spectrogram generation method is used that skips complicated waveform generation. In addition, the aperiodic components in speech are accurately predicted through differentiable DSP, enhancing the method's applicability to speech signal processing. Experimental results showed that the proposed method outperformed both baseline DSP and SSL-based pitch estimation methods, attributed to the effective integration of SSL and DSP.

Figures

Figures reproduced from arXiv: 2507.17208 by the authors.

Figure 1
Figure 1. Model architecture of SLASH. The different types of arrows (dashed, dotted, and solid) are used solely for visibility purposes. 2.2. Pitch consistency loss Following the success in previous works [7, 8, 12], we adopted the loss of pitch consistency. The key idea of this loss is to predict the pitch difference between the original and the pitch-shifted waveform to learn the relative pitch. For train￾ing, Constant-Q T… view at source ↗
Figure 2
Figure 2. An example of pitch guide G with log spectrogram. solute pitch, the pitch guide loss is defined as follows: Lg = 1 T XT t=1 max  1 − X f Pt,f · Gt,f − m, 0   , (3) where Pt,f represents P at time frame t and frequency f, and m is the hinge parameter to relax the constraint. Pitch guide loss is also applied to P shift. In this case, the pitch guide is shifted according to the amount of pitch shift ∆f: Lg-shift = … view at source ↗
Figure 3
Figure 3. An example of the target spectrum St and the pseudo periodic excitation spectrum E ∗ t generated from optimized pt. Then, a triangle wave oscillating X between −1 and 1 is gen￾erated based on Φt,k: Xt,k =  −1, if Φt,k < 0.5 4 |Φt,k − ⌊Φt,k⌋ − 0.5| − 1, otherwise, where ⌊·⌋ denotes the floor function. The pseudo periodic exci￾tation spectrogram E ∗ p is then obtained as: E ∗ p = max(X, ε) 2 + |Z · ε| , (4) where ε d… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: illustrates the RPA and log-F0 RMSE under conditions with added white noise. The results demonstrate that SLASH exhibits high robustness even in noisy environments. Notably, at an SNR of 0 dB, SLASH experienced minimal deterioration, while other systems showed signific…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 37 canonical work pages

  1. [1]

    It has been applied to many kinds of applica- tions, such as text-to-speech and emotion recognition, among others [1, 2, 3]

    Introduction Pitch estimation is a fundamental task in the field of speech sig- nal processing. It has been applied to many kinds of applica- tions, such as text-to-speech and emotion recognition, among others [1, 2, 3]. Historically, this task has been performed us- ing digital signal processing (DSP)-based methods [4, 5, 6], but with the advent of deep ...

    work page

  2. [2]

    SLASH: Self-Supervised Speech Pitch Estimation Leveraging DSP-derived Absolute Pitch

    SLASH 2.1. Model overview Figure 1 illustrates the architecture of SLASH. As shown in the figure, only the Pitch Encoder is a DNN module, while all other modules are composed of DSP. When viewed as a whole, SLASH functions as an analysis-synthesis system based on the source-filter model [16, 17, 18]. Specifically, the model decom- poses the waveform w int...

    work page Pith review arXiv 2025

  3. [3]

    Experimental setup Datasets: In our experiments, we used two datasets

    Experiments 3.1. Experimental setup Datasets: In our experiments, we used two datasets. The first one is LibriTTS-R [25], a high-quality multi-speaker English corpus of 585 hours of read English speech designed for text- to-speech use. The second one is MIR-1K [26], consisting of 1,000 song clips extracted from 110 karaoke songs. It contains both vocal an...

    work page

  4. [4]

    By incorporating the absolute pitch into the model, SLASH enhanced the pitch prediction accuracy of conventional SSL-based methods, which depend on relative pitch objectives

    Conclusion We proposed SLASH, an SSL-based pitch estimation method with DSP-derived absolute pitch information. By incorporating the absolute pitch into the model, SLASH enhanced the pitch prediction accuracy of conventional SSL-based methods, which depend on relative pitch objectives. Future work includes ex- panding SLASH to real-time pitch estimation

    work page

  5. [5]

    A survey on neural speech synthesis,

    X. Tan, T. Qin, F. Soong, and T.-Y . Liu, “A survey on neural speech synthesis,” arXiv preprint arXiv:2106.15561, 2021

    arXiv 2021

  6. [6]

    Speech emotion recognition using deep learning techniques: A review,

    R. A. Khalil, E. Jones, M. I. Babar, T. Jan, M. H. Zafar, and T. Alhussain, “Speech emotion recognition using deep learning techniques: A review,”IEEE Access, vol. 7, pp. 117 327–117 345, 2019

    work page 2019

  7. [7]

    An overview of voice conversion and its challenges: From statistical modeling to deep learning,

    B. Sisman, J. Yamagishi, S. King, and H. Li, “An overview of voice conversion and its challenges: From statistical modeling to deep learning,” IEEE/ACM Trans. on Audio, Speech, and Lang. Process., vol. 29, pp. 132–157, 2020

    work page 2020

  8. [8]

    PYIN: A fundamental frequency estimator using probabilistic threshold distributions,

    M. Mauch and S. Dixon, “PYIN: A fundamental frequency estimator using probabilistic threshold distributions,” in Proc. ICASSP, 2014, pp. 659–663

    work page 2014

Show all 39 references
  1. [9]

    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,” J. Acoust. Soc. Am. , vol. 124, no. 3, pp. 1638–1652, 2008

    2008

  2. [10]

    A robust algorithm for pitch tracking (RAPT),

    D. Talkin and W. B. Kleijn, “A robust algorithm for pitch tracking (RAPT),” Speech coding and synthesis , vol. 495, p. 518, 1995

    1995

  3. [11]

    PESTO: Pitch estimation with self-supervised transposition-equivariant objec- tive,

    A. Riou, S. Lattner, G. Hadjeres, and G. Peeters, “PESTO: Pitch estimation with self-supervised transposition-equivariant objec- tive,” in Proc. ISMIR, 2023

    2023

  4. [12]

    SPICE: Self-supervised pitch estimation,

    B. Gfeller, C. Frank, D. Roblek, M. Sharifi, M. Tagliasacchi, and M. Velimirovi ´c, “SPICE: Self-supervised pitch estimation,” IEEE/ACM Trans. on Audio, Speech, and Lang. Process. , vol. 28, pp. 1118–1128, 2020

    2020

  5. [13]

    Crepe: A convolu- tional representation for pitch estimation,

    J. W. Kim, J. Salamon, P. Li, and J. P. Bello, “Crepe: A convolu- tional representation for pitch estimation,” inProc. ICASSP, 2018, pp. 161–165

    2018

  6. [14]

    DeepF0: End-to-end fundamen- tal frequency estimation for music and speech signals,

    S. Singh, R. Wang, and Y . Qiu, “DeepF0: End-to-end fundamen- tal frequency estimation for music and speech signals,” in Proc. ICASSP, 2021, pp. 61–65

    2021

  7. [15]

    Noise-robust DSP-assisted neural pitch estimation with very low complexity,

    K. Subramani, J.-M. Valin, J. B ¨uthe, P. Smaragdis, and M. Good- win, “Noise-robust DSP-assisted neural pitch estimation with very low complexity,” inProc. ICASSP, 2024, pp. 11 851–11 855

    2024

  8. [16]

    NANSY++: Unified voice synthesis with neural analysis and synthesis,

    H.-S. Choi, J. Yang, J. Lee, and H. Kim, “NANSY++: Unified voice synthesis with neural analysis and synthesis,”arXiv preprint arXiv:2211.09407, 2022

    2022 arXiv

  9. [17]

    Singing voice separa- tion and vocal F0 estimation based on mutual combination of ro- bust principal component analysis and subharmonic summation,

    Y . Ikemiya, K. Itoyama, and K. Yoshii, “Singing voice separa- tion and vocal F0 estimation based on mutual combination of ro- bust principal component analysis and subharmonic summation,” IEEE/ACM Trans. on Audio, Speech, and Lang. Process. , vol. 24, no. 11, pp. 2084–2095, 2016

    2016

  10. [18]

    Unsupervised harmonic parameter estimation using differentiable DSP and spectral opti- mal transport,

    B. Torres, G. Peeters, and G. Richard, “Unsupervised harmonic parameter estimation using differentiable DSP and spectral opti- mal transport,” in Proc. ICASSP, 2024, pp. 1176–1180

    2024

  11. [19]

    DDSP: Differentiable digital signal processing,

    J. Engel, L. Hantrakul, C. Gu, and A. Roberts, “DDSP: Differentiable digital signal processing,” arXiv preprint arXiv:2001.04643, 2020

    2001 arXiv

  12. [20]

    Technical foundations of tandem- straight, a speech analysis, modification and synthesis frame- work,

    H. Kawahara and M. Morise, “Technical foundations of tandem- straight, a speech analysis, modification and synthesis frame- work,” Sadhana, vol. 36, pp. 713–727, 2011

    2011

  13. [21]

    WORLD: A vocoder- based high-quality speech synthesis system for real-time applica- tions,

    M. Morise, F. Yokomori, and K. Ozawa, “WORLD: A vocoder- based high-quality speech synthesis system for real-time applica- tions,” IEICE Trans. on Inf. and Syst. , vol. 99, no. 7, pp. 1877– 1884, 2016

    2016

  14. [22]

    A spectral envelope estimation method based on F0-adaptive multi-frame integration analysis,

    T. Nakano and M. Goto, “A spectral envelope estimation method based on F0-adaptive multi-frame integration analysis,” in Proc. SAPA-SCALE 2012, 2012, pp. 11–16

    2012

  15. [23]

    Differentiable world synthesizer-based neural vocoder with application to end-to-end audio style transfer,

    S. Nercessian, “Differentiable world synthesizer-based neural vocoder with application to end-to-end audio style transfer,”arXiv preprint arXiv:2208.07282, 2022

    2022 arXiv

  16. [24]

    Calculation of a constant Q spectral transform,

    J. C. Brown, “Calculation of a constant Q spectral transform,” J. Acous. Soc. Am., vol. 89, no. 1, pp. 425–434, 1991

    1991

  17. [25]

    Robust estimation of a location parameter,

    P. J. Huber, “Robust estimation of a location parameter,” Ann. Math. Statist., vol. 35, no. 1, pp. 73–101, 1964

    1964

  18. [26]

    Spectral smoothing technique in PARCOR speech analysis-synthesis,

    Y . Tohkura, F. Itakura, and S. Hashimoto, “Spectral smoothing technique in PARCOR speech analysis-synthesis,” IEEE Trans. on Acoustics, Speech, and Signal Process., vol. 26, no. 6, pp. 587– 596, 1978

    1978

  19. [27]

    Learning to learn by gradient descent by gradient descent,

    M. Andrychowicz, M. Denil, S. Gomez, M. W. Hoffman, D. Pfau, T. Schaul, B. Shillingford, and N. De Freitas, “Learning to learn by gradient descent by gradient descent,” in Proc. NIPS, vol. 29, 2016

    2016

  20. [28]

    A spectral energy distance for parallel speech synthesis,

    A. Gritsenko, T. Salimans, R. van den Berg, J. Snoek, and N. Kalchbrenner, “A spectral energy distance for parallel speech synthesis,” in Proc. NIPS, vol. 33, 2020, pp. 13 062–13 072

    2020

  21. [29]

    LibriTTS-R: A restored multi-speaker text-to-speech corpus,

    Y . Koizumi, H. Zen, S. Karita, Y . Ding, K. Yatabe, N. Morioka, M. Bacchiani, Y . Zhang, W. Han, and A. Bapna, “LibriTTS-R: A restored multi-speaker text-to-speech corpus,” in Proc. Inter- speech, 2023, pp. 5496–5500

    2023

  22. [30]

    On the improvement of singing voice separation for monaural recordings using the MIR-1K dataset,

    C.-L. Hsu and J.-S. R. Jang, “On the improvement of singing voice separation for monaural recordings using the MIR-1K dataset,” IEEE Trans. on Audio, Speech, and Lang. Process., vol. 18, no. 2, pp. 310–319, 2009

    2009

  23. [31]

    ESPnet-TTS: Uni- fied, reproducible, and integratable open source end-to-end text- to-speech toolkit,

    T. Hayashi, R. Yamamoto, K. Inoue, T. Yoshimura, S. Watanabe, T. Toda, K. Takeda, Y . Zhang, and X. Tan, “ESPnet-TTS: Uni- fied, reproducible, and integratable open source end-to-end text- to-speech toolkit,” in Proc. ICASSP, 2020, pp. 7654–7658

    2020

  24. [32]

    Decoupled weight decay regulariza- tion,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regulariza- tion,” in Proc. ICLR, 2019

    2019

  25. [33]

    Fast and reliable F0 estimation method based on the period extraction of vocal fold vibration of singing voice and speech,

    M. Morise, H. Kawahara, and H. Katayose, “Fast and reliable F0 estimation method based on the period extraction of vocal fold vibration of singing voice and speech,” in Proc. AES 35th Inter- national Conference, 2009

    2009

  26. [34]

    Harvest: A high-performance fundamental frequency estimator from speech signals,

    M. Morise et al. , “Harvest: A high-performance fundamental frequency estimator from speech signals,” in Proc. Interspeech, 2017, pp. 2321–2325

    2017

  27. [35]

    Multiple fundamental fre- quency estimation by modeling spectral peaks and non-peak re- gions,

    Z. Duan, B. Pardo, and C. Zhang, “Multiple fundamental fre- quency estimation by modeling spectral peaks and non-peak re- gions,” IEEE Trans. on Audio, Speech, and Lang. Process. , vol. 18, no. 8, pp. 2121–2133, 2010

    2010

  28. [36]

    MedleyDB: A multitrack dataset for annotation- intensive MIR research

    R. M. Bittner, J. Salamon, M. Tierney, M. Mauch, C. Cannam, and J. P. Bello, “MedleyDB: A multitrack dataset for annotation- intensive MIR research.” in Proc. ISMIR, vol. 14, 2014, pp. 155– 160

    2014

  29. [37]

    An analysis/synthesis framework for automatic F0 annotation of multitrack datasets,

    J. Salamon, R. M. Bittner, J. Bonada, J. J. Bosch, E. G ´omez Guti ´errez, and J. P. Bello, “An analysis/synthesis framework for automatic F0 annotation of multitrack datasets,” in Proc. ISMIR, 2017

    2017

  30. [38]

    Neural audio synthesis of musi- cal notes with wavenet autoencoders,

    J. Engel, C. Resnick, A. Roberts, S. Dieleman, M. Norouzi, D. Eck, and K. Simonyan, “Neural audio synthesis of musi- cal notes with wavenet autoencoders,” in Proc. ICML, 2017, pp. 1068–1077

    2017

  31. [39]

    Melody extraction from polyphonic music signals: Approaches, applica- tions, and challenges,

    J. Salamon, E. G ´omez, D. P. Ellis, and G. Richard, “Melody extraction from polyphonic music signals: Approaches, applica- tions, and challenges,”IEEE Signal Processing Magazine, vol. 31, no. 2, pp. 118–134, 2014

    2014

Pith tools

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