Pith. sign in

REVIEW 4 major objections 6 minor 20 references

A Robust Method for Pitch Tracking in the Frequency Following Response using Harmonic Amplitude Summation Filterbank

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A stimulus-aware harmonic filterbank estimates F0 contours in the Frequency Following Response more accurately than autocorrelation, cutting average RMSE by 8.8–47.4%.

desk verdict A useful adaptation of harmonic summation to FFR pitch tracking, but the headline RMSE gains over ACF are inflated by post hoc selection of the harmonic count K and an un-matched ACF baseline. read the letter →

arxiv 2506.19253 v1 pith:LCOOOUNF submitted 2025-06-24 cs.SD eess.ASeess.SP

Sajad Sadeghkhani , Maryam Karimi Boroujeni , Hilmi R. Dajani , Saeid R. Seydnejad , Christian Giguère This is my paper · ORCID
classification cs.SDeess.ASeess.SP
keywords FrequencyFollowingResponsefundamentaltrackingharmonicamplitudesummationpitchauditorybrainstemprominencepeakpickingautocorrelationfunction
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

The paper claims that pitch tracking in the Frequency Following Response (FFR) — the brain's scalp-recorded response to sound — is more accurate when the algorithm knows the stimulus pitch in advance and searches only near it. It introduces the Harmonic Amplitude Summation (HAS) filterbank, which sums amplitudes at candidate F0 and its first few harmonics while suppressing non-harmonic frequencies, and combines it with prominence-based peak picking (HAS-PR). On FFRs from 16 normal-hearing listeners responding to four natural-speech "balloon" stimuli with F0 ranging from 89 Hz to 452 Hz, the paper reports that HAS-PR lowers root-mean-square error between stimulus and response F0 contours by 8.8% to 47.4% relative to the standard autocorrelation method, with larger gains for higher-pitched female voices. If correct, this gives auditory researchers a more reliable way to trace how the brainstem tracks pitch in natural speech, with potential value for clinical screening.

What carries the argument

The central object is the Harmonic Amplitude Summation (HAS) filterbank: a bank of synthetic periodic filters, one for each F0 candidate at 1 Hz resolution between 80 Hz and 500 Hz, each containing a fundamental plus K harmonics (optimally K=4 for the low-F0 male speech and K=2 for the high-F0 female speech). Each filter's magnitude DFT is multiplied pointwise with the response frame's magnitude DFT and summed, so a candidate F0 scores high only when energy lines up at its fundamental and its harmonics. The second mechanism is prominence-based peak picking: instead of taking the tallest peak in the score curve, the algorithm takes the peak that stands out most against its local surroundings, which makes the estimate insensitive to the overall spectral downward slope produced by noise. The F0 search window is centered on the known stimulus F0 of the same aligned frame, with a ±50 Hz threshold, which prevents octave errors and lets the method keep frequency resolution.

What would settle it

Simulate a synthetic FFR frame exactly as in Eq. (1) with a true F0 set 60 Hz above the stimulus F0 and K=2 harmonics added, then run HAS-PR and check whether the selected peak is near the true F0 or falls back to an alias inside the ±50 Hz window; a correct estimate outside the window would refute the method's search constraint as load-bearing.

Watch

Extended reading notes

Core claim

The central claim is that harmonic structure, not just periodicity, carries reliable pitch information in FFRs, and that a filterbank can exploit it once the search is constrained by the known stimulus F0. The authors argue that previous speech and music pitch trackers fail on FFRs because they search too wide a range, weight harmonics logarithmically, include too many harmonics, and pick the highest rather than the most prominent peak. HAS-PR addresses each issue: it uses a DFT on a linear scale, aggregates only the first 2–4 harmonics, restricts candidate F0s to a 50 Hz window around the stimulus F0 of the time-aligned frame, and selects the most prominent peak in the summed spectrum. In their recordings, this reduces RMSE relative to ACF by 8.8% for Male Sad, 31.2% for Male Happy, 37.8% for Female Sad, and 47.4% for Female Happy stimuli, and also lowers gross pitch error in all conditions. The method is also reported to have lower RMSE than Praat, PEFAC, HarmoF0, and a Bayesian pitch tracker in all four stimulus conditions.

Load-bearing premise

The method assumes the neural F0 of the response never wanders more than 50 Hz away from the stimulus F0; the reported accuracy only covers frames where that window contains the true response pitch, and the paper does not test what happens when this assumption is violated.

Editorial extensions

If this is right

  • If HAS-PR works as reported, researchers can extract F0 contours from FFRs evoked by natural speech with lower error than ACF, especially for high-pitched female talkers where the response harmonic structure is weaker.
  • Because the method uses only the first few harmonics and a constrained search, it is designed to avoid octave errors, a failure mode that the paper reports affects several speech-based baseline methods.
  • The method's reliance on the known stimulus F0 means it can be applied to any FFR paradigm in which the stimulus F0 contour is available, not just the four emotional speech stimuli tested here.
  • The paper's sweep-count comparisons indicate that the advantage of HAS-PR over ACF persists as the number of averaged response sweeps changes, suggesting the method could support shorter or more flexible recording protocols.
  • Since HAS-PR operates on a per-frame DFT, it can be combined with existing FFR preprocessing such as band-pass filtering and coherent averaging, making it a drop-in replacement for autocorrelation in current pipelines.

Reading between the lines

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

  • If the response F0 ever deviates by more than 50 Hz from the stimulus F0, HAS-PR will be blind to it; testing this boundary by comparing HAS-PR with a broad-range pitch tracker on recordings where the pitch percept is shifted (for example, via a missing fundamental or masking) would directly probe the method's main assumption.
  • The prominence-based peak selector could transfer to speech pitch trackers as a detrending-free way to handle spectral tilt, an extension the paper only hints at when discussing why the highest peak is unreliable for FFRs.
  • The condition-dependent optimal harmonic count (K=4 for low-F0 stimuli, K=2 for high-F0 stimuli) suggests an adaptive harmonic-count rule might improve accuracy across speakers and pitches, since the paper fixes K per condition rather than adapting it frame by frame.
  • Because the method needs only the stimulus F0 contour and a DFT, it could be implemented in near-real-time for neurofeedback or closed-loop auditory experiments, although the paper does not address real-time operation.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes HAS-PR, a frequency-domain pitch-tracking method for the Frequency Following Response (FFR). The method builds a filterbank of harmonic templates (Eq. 3), aggregates harmonic amplitudes (Eq. 7), and selects the most prominent peak within a ±50 Hz window centered on the known stimulus F0 (Section 2.4). The authors evaluate HAS-PR against ACF and several speech-pitch algorithms on FFRs from 16 participants to four natural speech stimuli, and report that HAS-PR reduces mean RMSE relative to ACF by 8.8% to 47.4% depending on stimulus. They also report Gross Pitch Error and RMSE20 metrics, and analyze performance versus number of averaged sweeps.

Significance. If the reported gains are unbiased, the method could be a practically useful advance for FFR F0 tracking, which is currently dominated by ACF. The paper is one of the first to explicitly adapt harmonic-structure-based pitch tracking to FFRs, and the use of prominence-based peak picking to counter spectral tilt is a sensible idea. The inclusion of GPE and RMSE20 metrics, along with the sweep-count analysis, is a strength. However, the significance hinges on whether the comparison to ACF is fair and whether the reported improvements are in-sample artifacts of parameter selection.

major comments (4)
  1. [§3.2, Table 1 and Eq. (3)] The number of harmonics K was selected post hoc on the same test data used for the performance evaluation. The text in Section 3.2 states that the optimal K 'resulted in the minimum RMSE in Table 1' (K=4 for Male Sad, K=2 for the others). This makes the reported RMSE values in-sample optima rather than unbiased performance estimates. The paired t-tests in Section 3.2 comparing HAS-PR with the optimal K against ACF or K=1 are therefore also post hoc. The central claim that HAS-PR 'outperformed' ACF is not established by this analysis. Please re-run the evaluation with a fixed K chosen a priori, or with K selected by cross-validation on a development subset, and report the corresponding RMSE values and test statistics.
  2. [§2.4 and Table 1] The ACF baseline is described only as 'the Autocorrelation Function' and the implementation details are not given. In particular, it is not stated whether ACF was given the same ±50 Hz prior centered on the stimulus F0 that HAS-PR uses, nor whether the same frame length, step size, and band-pass filtering were applied. If ACF was not given the stimulus-F0 prior, part of the reported RMSE reduction could derive from the prior rather than from harmonic summation or prominence peak picking. Please specify the ACF implementation completely, and also report an ACF variant that uses the same F0 search window so that the comparison isolates the algorithmic contribution.
  3. [§2.4 and §3.1] The evaluation may be partly circular because the error metric is the difference between the estimated response F0 and the known stimulus F0, while the algorithm restricts its search to a ±50 Hz interval centered on that same stimulus F0. This design guarantees that any estimate is within 50 Hz of the reference, and it mixes the effect of a strong prior with the effect of the harmonic-sum filterbank. The paper should quantify how much of the error reduction is due to the search window itself. A concrete test: compare HAS-PR with a version of ACF that is given the same 50 Hz window, and also report results when the window is widened (e.g., ±100 Hz or a full 80-500 Hz search) to show that the algorithm still tracks accurately without the tight prior.
  4. [§3.2] The statistical analysis is underpowered for the strength of the claims. Paired t-tests are performed on the same data used to select K, and multiple comparisons (three sets of comparisons across four stimuli) are made without correction. The p-value of 0.078 for the FS condition in the K vs. K=1 comparison is reported as non-significant, but the text does not discuss whether the pattern of results would survive a multiple-comparison correction. Please provide a corrected analysis or clearly state the number of comparisons and the correction procedure.
minor comments (6)
  1. [§2.4] The definition of 'prominence' is not given explicitly; the text only cites Kirmse & de Ferranti (2017). Please specify the algorithm used for prominence computation, as the exact definition affects reproducibility.
  2. [§2.4, Eq. (5)-(7)] The matrix H is defined via the DFT of each filter, but the size of H and the normalization of the DFT are not stated. Please clarify how the filters are normalized and whether the summation in Eq. (7) is over the positive-frequency bins only.
  3. [§2.2] The stimulus F0 extraction uses the median of several MATLAB pitch methods with manual correction of octave errors. This is a reasonable approach, but the manual correction step is subjective; please state how many frames were corrected and whether the correction was blinded to the subsequent FFR analysis.
  4. [Table 1] The table reports means across participants but not standard deviations, even though the text says 'mean and standard deviation' are presented. Please add the SD values or remove the claim.
  5. [§3.1] In Figure 3, the performance is shown versus the number of averaged sweeps, but it is unclear whether the number of sweeps was varied per participant and whether the results are averaged over participants or shown as medians. Please clarify the aggregation procedure.
  6. [Abstract and §4] The phrase 'within each response and stimulus F0 contour pair' in the abstract is awkward; consider rephrasing to 'across pairs of response and stimulus F0 contours.'

Circularity Check

2 steps flagged · score 6.0 of 10

Reported RMSE gains are partly in-sample: K is tuned on the test table and the F0 search is centered on the stimulus F0 used as the error reference.

  1. fitted input called prediction [Section 3.2 and Table 1; Discussion (improvement percentages)]
    "The optimal K value in Eq. (3) which resulted in the minimum RMSE in Table 1 was higher for the MS condition, which had a lower F0 contour. This observation indicates that higher stimulus F0 contours lead to weaker response harmonic structures. Specifically, the optimal K for the MS condition was 4, while it was 2 for the rest."

    K is chosen per stimulus by minimizing the very RMSE values that are then reported as the method's performance in Table 1 and in the Abstract/Discussion (8.8-47.4% improvement over ACF). The paired t-tests in Section 3.2 compare HAS-PR with this optimally tuned K against ACF and against K=1, so the claimed advantage reflects post hoc parameter selection on the test data rather than an unbiased evaluation of a fixed algorithm. The headline RMSE reductions are therefore in-sample optima, not predictions of how HAS-PR would perform with a pre-specified K.

  2. self definitional [Abstract; Section 2.4 (F0 Tracking Algorithm)]
    "we introduce a stimulus-aware filterbank that selectively aggregates amplitudes at F0 and its harmonics ... evaluates F0 candidates only within a range centered around the stimulus F0. ... The objective is to identify F0_i at which y in Eq. (7) produces the most prominent peak ... within a defined threshold (50 Hz) around the stimulus F0 of the same time-aligned frame."

    The response F0 estimate is constrained to lie within 50 Hz of the stimulus F0, and the evaluation metric is the RMSE between the estimated response F0 and that same stimulus F0 contour. By construction the estimator cannot make errors larger than the 50 Hz window, while the ACF baseline is not stated to receive the same prior and can commit octave errors. The reported superiority of HAS-PR over ACF is thus partly an artifact of giving HAS-PR the reference value in advance, not a pure test of harmonic summation or prominence peak-picking.

full rationale

The paper is self-contained and does not rely on a load-bearing self-citation chain: the method is implemented from first principles (DFT filterbank, harmonic summation, prominence peak-picking) and the data are from the authors' own prior recording study, which is normal rather than circular. However, the central quantitative claim is partially circular. First, the number of harmonics K is selected per stimulus as the value that minimizes the RMSE in Table 1, and the same Table 1 values are then used to claim 8.8-47.4% RMSE reductions over ACF; those numbers are in-sample optima rather than unbiased estimates for a fixed model. Second, the algorithm searches for F0 only within a 50 Hz window centered on the stimulus F0, which is exactly the reference contour used to compute the reported RMSE; this makes large errors impossible by construction and confounds the comparison with methods that do not receive the same prior. These issues are explicit in the text and do not require speculation, but they do not reduce the method to a trivial identity: within the window, the selected prominent harmonic peak is still computed from the FFR spectrum, and the method may still have genuine value. The score of 6 reflects two load-bearing 'predictions' that are partly built from the evaluation target and the test data.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the stimulus-aware search window, the harmonic model, and the accuracy of the stimulus F0 reference; none of these are tested or justified with external data.

free parameters (2)
  • Number of harmonics K = 4 for Male Sad, 2 for the other three stimuli
    Chosen per stimulus to minimize the RMSE in Table 1, i.e., fitted to the test data.
  • F0 search window width around stimulus F0 = 50 Hz
    Hand-chosen threshold; not varied or justified based on FFR physiology; defines the range of F0 candidates.
assumptions (4)
  • domain assumption The FFR F0 remains within 50 Hz of the stimulus F0
    Used to restrict the F0 candidate range in the filterbank; if violated, the method cannot track the true response F0.
  • domain assumption The FFR can be modeled as a harmonic series with K components plus noise (Eq. 1)
    The filterbank is constructed from harmonic templates; the validity of the harmonic model for FFRs is assumed.
  • domain assumption The stimulus F0 contour obtained from MATLAB pitch commands and manual correction is accurate and can serve as ground truth
    The evaluation metric compares the estimated response F0 to the estimated stimulus F0; errors in the stimulus reference directly affect reported RMSE.
  • domain assumption Phase-locking in auditory neurons is too weak above 2000 Hz
    Used to set the upper cutoff of the band-pass filter for the response; affects which frequencies are available for harmonic summation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Robust Method for Pitch Tracking in the Frequency Following Response using Harmonic Amplitude Summation Filterbank." pith.science (2026). https://pith.science/paper/LCOOOUNF

@misc{pith2026250619253,
  author       = {Pith},
  title        = {Pith review of: A Robust Method for Pitch Tracking in the Frequency Following Response using Harmonic Amplitude Summation Filterbank},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LCOOOUNF}},
  note         = {Machine review of arXiv:2506.19253}
}
read the original abstract

The Frequency Following Response (FFR) reflects the brain's neural encoding of auditory stimuli including speech. Because the fundamental frequency (F0), a physical correlate of pitch, is one of the essential features of speech, there has been particular interest in characterizing the FFR at F0, especially when F0 varies over time. The standard method for extracting F0 in FFRs has been the Autocorrelation Function (ACF). This paper investigates harmonic-structure-based F0 estimation algorithms, originally developed for speech and music, and resolves their poor performance when applied to FFRs in two steps. Firstly, given that unlike in speech or music, stimulus F0 of FFRs is already known, we introduce a stimulus-aware filterbank that selectively aggregates amplitudes at F0 and its harmonics while suppressing noise at non-harmonic frequencies. This method, called Harmonic Amplitude Summation (HAS), evaluates F0 candidates only within a range centered around the stimulus F0. Secondly, unlike other pitch tracking methods that select the highest peak, our method chooses the most prominent one, as it better reflects the underlying periodicity of FFRs. To the best of our knowledge, this is the first study to propose an F0 estimation algorithm for FFRs that relies on harmonic structure. Analyzing recorded FFRs from 16 normal hearing subjects to 4 natural speech stimuli with a wide F0 variation from 89 Hz to 452 Hz showed that this method outperformed ACF by reducing the average Root-Mean-Square-Error (RMSE) within each response and stimulus F0 contour pair by 8.8% to 47.4%, depending on the stimulus.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 17 canonical work pages

  1. [1]

    L., MacDonald, E

    Bachmann, F. L., MacDonald, E. N., & Hjortkjær, J. (2021). Neural measures of pitch processing in EEG responses to running speech. Frontiers in Neuroscience, 15, 738408, https://doi.org/10.3389/fnins.2021.738408

    work page arXiv 2021

  2. [2]

    Boersma, P., & Van Heuven, V . (2001). Speak and unSpeak with PRAAT. Glot International, 5(9/10), 341-347,

    work page 2001

  3. [3]

    Camacho, A., & Harris, J. G. (2008). A sawtooth waveform inspired pitch estimator for speech and music. The Journal of the Acoustical Society of America, 124(3), 1638-1652, https://doi.org/10.1121/1.2951592

    work page doi:10.1121/1.2951592 2008

  4. [4]

    R., Purcell, D., Wong, W., Kunov, H., & Picton, T

    Dajani, H. R., Purcell, D., Wong, W., Kunov, H., & Picton, T. W. (2005). Recording human evoked potentials that follow the pitch contour of a natural vowel. IEEE Transactions on Biomedical Engineering, 52(9), 1614-1618, https://doi.org/10.1109/TBME.2005.851499 De Medeiros, B. R., Cabral, J. P., Meireles, A. R., & Baceti, A. A. (2021). A comparative study ...

    work page arXiv 2005

  5. [5]

    Drugman, T., & Alwan, A. (2019). Joint robust voicing detection and pitch estimation based on residual harmonics. arXiv preprint arXiv:2001.00459, https://doi.org/10.48550/arXiv.2001.00459

    work page Pith review arXiv doi:10.48550/arxiv.2001.00459 2019

  6. [6]

    E., Etard, O., & Reichenbach, T

    Forte, A. E., Etard, O., & Reichenbach, T. (2017). The human auditory brainstem response to running speech reveals a subcortical mechanism for selective attention. Elife, 6, e27203, https://doi.org/10.7554/eLife.27203

    work page doi:10.7554/elife.27203 2017

  7. [7]

    Gonzalez, S., & Brookes, M. (2014). PEFAC-A pitch estimation algorithm robust to high levels of noise. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 22(2), 518-530, https://doi.org/10.1109/TASLP.2013.2295918

    work page arXiv 2014

  8. [8]

    Jadoul, Y ., Thompson, B., & De Boer, B. (2018). Introducing parselmouth: A python interface to praat. Journal of Phonetics, 71, 1-15,

    work page 2018

Show all 20 references
  1. [9]

    Wu, G., & Lin, C.-D. (2011). Cross-linguistic comparison of frequency-following responses to voice pitch in American and Chinese neonates and adults. Ear and hearing, 32(6), 699-707, https://doi.org/10.1097/aud.0b013e31821cc0df Karimi Boroujeni, M., Sadeghkhani, S., Seyednejad...

    2011 doi

  2. [10]

    Kirmse, A., & de Ferranti, J. (2017). Calculating the prominence and isolation of every mountain in the world. Progress in Physical Geography, 41(6), 788-802,

    2017

  3. [11]

    T., & Cariani, P

    Krishnan, A., Xu, Y ., Gandour, J. T., & Cariani, P. A. (2004). Human frequency-following response: representation of pitch contours in Chinese tones. Hearing Research, 189(1-2), 1- 12, https://doi.org/10.1016/s0378-5955(03)00402-7

    2004 doi

  4. [12]

    Krizman, J., & Kraus, N. (2019). Analyzing the FFR: A tutorial for decoding the richness of auditory function. Hearing Research, 382, 107779, https://doi.org/10.1016/j.heares.2019.107779

    2019

  5. [13]

    Li, Q., Millard, K., Tetnowski, J., Narayana, S., & Cannito, M. (2023). Acoustic analysis of intonation in persons with Parkinson's disease receiving transcranial magnetic stimulation and intensive voice treatment. Journal of Voice, 37(2), 203-214, https://doi.org/10.1016/j.jv...

    2023 doi

  6. [14]

    K., & Lee, A

    Maddox, R. K., & Lee, A. K. (2018). Auditory brainstem responses to continuous natural speech in human listeners. Eneuro, 5(1), https://doi.org/10.1523/eneuro.0441-17.2018

    2018 doi

  7. [15]

    Rabiner, L. (1977). On the use of autocorrelation analysis for pitch detection. IEEE Transactions on Acoustics, Speech, and Signal Processing, 25(1), 24-33, https://doi.org/10.1109/TASSP.1977.1162905

    1977

  8. [16]

    K., Jensen, J

    Shi, L., Nielsen, J. K., Jensen, J. R., Little, M. A., & Christensen, M. G. (2019). Robust Bayesian pitch tracking based on the harmonic model. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 27(11), 1737-1751, 10.1109/TASLP.2019.2930917

    2019

  9. [17]

    Singh, L., & Fu, C. S. (2016). A new view of language development: the acquisition of lexical tone. Child Development, 87(3), 834-854, https://doi.org/10.1111/cdev.12512

    2016 doi

  10. [18]

    Wei, W., Li, P., Yu, Y ., & Li, W. (2022). Harmof0: Logarithmic scale dilated convolution for pitch estimation. 2022 IEEE International Conference on Multimedia and Expo (ICME), https://doi.org/10.1109/ICME52920.2022.9858935

    2022

  11. [19]

    Xu, Q., & Ye, D. (2014). Evaluation of a posteriori Wiener filtering applied to frequency-following response extraction in the auditory brainstem. Biomedical Signal Processing and Control, 14, 206-216, https://doi.org/10.1016/j.bspc.2014.08.003

    2014 doi

  12. [20]

    Zhou, K., Sisman, B., Liu, R., & Li, H. (2022). Emotional voice conversion: Theory, databases and esd. Speech Communication, 137, 1-18, https://doi.org/10.1016/j.specom.2021.11.006

    2022 doi

Pith tools

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