Audio Cross Verification Using Dual Alignment Likelihood Ratio Test

Arm Wonghirundacha; Heidi Lei; Irmak Bukey; TJ Tsai

REVIEW 3 major objections 3 minor 35 references

Audio Cross Verification Using Dual Alignment Likelihood Ratio Test

T0 review · 3 major / 3 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read This paper claims that audio tampering can be caught by comparing two explicit alignments of a query clip against a trusted reference, using a likelihood ratio to decide between clean and tampered hypotheses.

desk verdict The dual-alignment idea is useful and the empirical gains over the baseline look real, but the paper's central claim that it performs a likelihood ratio test is undercut by a mis-specified H2 model for insertions/deletions. read the letter →

arxiv 2607.18190 v1 pith:KODR5ZV4 submitted 2026-07-20 cs.SD

Heidi Lei , Arm Wonghirundacha , Irmak Bukey , TJ Tsai This is my paper · ORCID

classification cs.SD

keywords audiocrossverificationtamperingdetectionlikelihoodratiotesthiddenstatetimewarpingMFCCforensicsdualalignmentexplainability

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

This paper tries to establish that a short audio clip can be positively verified against a trusted reference recording by asking not whether the clip shows signs of tampering, but whether it matches the reference. It proposes a method called DA-LRT that computes the most likely alignment under two hypotheses — clean and tampered — and compares the likelihoods. The central claim is that this dual-alignment likelihood ratio detects inserted, deleted, and replaced audio with error rates near zero for long tampering and as low as 0.20% for one-quarter-second insertions, far outperforming a simple MFCC-distance baseline. The method is also fast (about 6 seconds total) and explains its decisions through the two explicit hypotheses.

What carries the argument

The central object is the Dual Alignment Likelihood Ratio Test (DA-LRT). It first preselects a candidate region using binary audio hashprints, then computes two alignments: a diagonal alignment under the clean hypothesis and a Hidden State Time Warping alignment under the tampered hypothesis, which permits (0,1) and (1,0) transitions to represent insertion and deletion. The third stage partitions query frames by whether the two alignments differ, models the matching-region feature differences as per-MFCC Gaussians, derives an independence-based non-matching model from those same marginals, and scores the query with a log likelihood-ratio statistic. The work the machinery does is to turn a bi

What would settle it

Construct tampered queries where the inserted or replaced segment is drawn from a different speaker, a different acoustic environment, or non-speech material, and measure the EER; if the independence-plus-marginal model is correct, detection should be at least as good as with same-source filler, but if the model is misspecified, EER will rise sharply, especially for short replacements.

Watch

Extended reading notes

Core claim

The paper's central claim is that audio cross verification is best framed as a choice between two explicit hypotheses about the query-to-reference relationship, rather than as a search for tampering artifacts. Under H1, the query is a faithful segment of the reference, so the alignment is a single diagonal offset. Under H2, the query may contain insertions, deletions, or replacements, so the alignment is allowed to drift, with a state-based dynamic program (Hidden State Time Warping) finding the most likely path. The method then splits the query frames into those where the two alignments agree (treated as reliable matches) and those where they differ (suspicious), models per-feature differen

Load-bearing premise

The tampered-region likelihood model assumes that in tampered regions the query frames are statistically independent of the reference frames and follow the same marginal feature distributions estimated from the clearly matching region; if real inserted or replaced material has different statistics — a different speaker, music, room acoustics, or any correlation with the reference — the H2 model is misspecified and the reported error rates may not hold.

Editorial extensions

If this is right

Social media platforms could attach 'verified against news source' labels to clips that match a trusted reference, giving viewers immediate confidence in viral video authenticity.
Cross verification offers a stable alternative to artifact-based tampering detection, since it does not depend on the sophistication of the tampering or anti-forensic countermeasures.
Because the method detects insertion and deletion by the resulting global desynchronization, even extremely short (0.25-second) tampering is caught reliably.
The likelihood ratio provides a tampering score suitable for setting decision thresholds, enabling receiver-operating-characteristic evaluation across application-specific trade-offs.
The approach can complement existing forensic tools, providing positive verification when a trusted reference exists and flagging when no match is found.

Reading between the lines

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

The independence assumption for tampered regions is a potential weak point; real tampering with foreign material (different speaker, music, noise) may break it, and the method's robustness to such mismatches is untested in this paper.
Because the tampered-region likelihood is derived from matching-region marginals, the method implicitly assumes that tampered content has the same overall feature statistics as the host recording; a testable extension is to measure how EER degrades when the inserted material comes from a different acoustic environment.
The runtime bottleneck is MFCC computation on the reference (5.82 seconds), so pre-computing features for known references could make verification near-real-time and scale to large reference archives.
The same dual-alignment logic could be extended to joint audio-visual verification, aligning a video query against a trusted broadcast and treating visual insertions as the analog of audio tampering.

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

I read DA-LRT carefully, and the short version is: the core idea is worth engaging with, but the statistical framing is overstated. What's new here is the dual-alignment setup: one alignment assumes a fixed offset, the other allows tampering via HSTW, and the two alignments are compared by a likelihood-ratio-like score. That is a reasonable and genuinely content-based approach to cross verification, distinct from ENF or artifact-based methods. The empirical results on DAPS show large EER improvements over the MFCC-Euclidean baseline, and the runtime is fine. The authors also shipped code, which is good practice.

The soft spot is not the absence of real-world validation; it's that Eq. 1–2 are not a valid log-likelihood ratio for the two stated hypotheses. For insertion or deletion, every query frame after the tampering point falls into partition B, but under the H2 alignment those frames are correctly matched to the reference—the insertion/deletion only changes the offset. The paper still scores those frames with the 'non-matching' independent-Gaussian model estimated from partition A. So H2 is evaluated under a model that is false for the majority of the observations it is applied to. That means the tampering score is a heuristic, not a likelihood ratio. The explainability claim in §4 is therefore weaker than advertised.

Other issues are more conventional: α, γ, β are never stated, no error bars are given, and the evaluation uses a single dataset with fillers drawn from the same recording, so real-world tampering with different speakers or acoustics is untested. The replacement-only results are the honest measure, and they're fine for L ≥ 0.5 s, so the method is not empty. But the zero EERs on insertion/deletion come from the global shift, which the heuristic detects easily—that part is believable.

Who is this for? People working on audio authentication, forensic verification, or social-media provenance tools. The paper deserves a serious referee because the problem is timely and the approach is novel despite the statistical flaw. A good referee should ask for hyperparameter values, error bars or a second dataset, and an honest revision that either fixes the likelihood calculation or explicitly describes the score as a heuristic. I would not cite it as-is, but I'd keep it on the reading list for the idea.

Referee Report

3 major / 3 minor

Summary. The paper proposes DA-LRT, a method for verifying whether a short audio query is taken without tampering from a trusted reference recording. It first preselects the approximate offset using hashprints, then computes two alignments between query and reference MFCC features: a diagonal alignment under a non-tampered hypothesis (H1) and an HSTW alignment with matching/tampered states under a tampered hypothesis (H2). The query frames are partitioned according to whether the two alignments agree; difference statistics are estimated from the agreeing partition and a log-likelihood-ratio score is computed on the disagreeing frames. The method is evaluated on DAPS clean recordings with simulated insertion, deletion, and replacement tampering at several bitrates and tampering durations. The reported EERs are much lower than those of the MFCC-Euclidean baseline, especially for short tampering, and the full test takes about 6.4 seconds per query.

Significance. If the method worked as described, it would be a practically useful and explainable complement to artifact-based tampering detection: it is fast, it positively verifies against a trusted source, and its dual-alignment design gives a natural way to localize tampering. The paper ships code, evaluates on a standard dataset, and includes a baseline comparison. The empirical results are striking, with near-zero EERs for insertion and deletion even at 0.25 s tampering. However, the central formal claim—that DA-LRT performs a likelihood ratio test on the two alignments—is not supported as written: the H2 likelihood used in Eq. (1) is mis-specified for the insertion/deletion cases that drive the reported performance. The method may still work as a heuristic, but the paper's statistical framing and explainability claims need substantial revision before the contribution can be accepted.

major comments (3)

[§2.3, Eq. (1)–(2)] The claimed likelihood-ratio test does not compute the likelihood of the stated hypotheses. For insertion and deletion, every query frame after the tampering point has H1 and H2 reference indices that differ by the inserted/deleted length, so those frames are assigned to partition B. Under H2, however, those frames are in the matching state: HSTW has shifted the offset and then resumed (1,1) transitions, so H2 aligns them correctly to the reference. Step 3 nevertheless scores all B frames with the non-matching model N(˜µ_f, ˜σ²_f) estimated from A. Thus Eq. (2) compares H1's wrong alignments under the matching model with H2's correct alignments under the non-matching model. The test statistic is therefore a heuristic score, not a log-likelihood ratio, and the explainability claim in §4 ('compares the likelihood of two interpretable hypotheses') is not justified. The fix is to condition t
[§2.3, third step] The construction of the non-matching distribution N(˜µ_f, ˜σ²_f) is not derived from the stated independence assumption. The text says that if query and reference features are independent draws from N(µ_f, σ²_f), then the difference distribution has mean and variance as given. For independent draws from the same Gaussian, the difference has mean 0 and variance 2σ²_f—not the displayed µ_f and the expression involving µquery_f and µref_f, which are never defined. Also, tilde µ_f is defined as the A-frame mean difference, which under independence would be µ_query − µ_ref and is not identifiable from paired A differences alone. As written, Eq. (1) is not a well-defined likelihood for H2. Please define the marginal statistics explicitly and derive the difference distribution, or clearly state the scoring rule as heuristic.
[§2.2, DP rules] The HSTW recurrence depends on hyperparameters α, γ, and β, but no numerical values are given anywhere in the paper. The train/test split in §3 is described, but no training or hyperparameter-selection procedure is reported. Since Table 1 reports EERs as low as 0.00 in several cells, the reader cannot tell whether the results are robust to these penalties or tuned to the test set. Please report the exact values used and include a sensitivity analysis over reasonable ranges.

minor comments (3)

[§3, data generation] The tampering simulation draws filler segments from the same compressed recording, so inserted/replaced content is perfectly matched in speaker, speaking style, and overall acoustic environment. This makes the detection task harder in one sense, but it also means the non-matching model is never tested against genuinely foreign material (different speaker, room acoustics, non-speech content). Please state this limitation explicitly in the conclusions.
[Table 1] No confidence intervals or variance estimates are reported for the EERs. With 2000 test queries per benchmark, a 0.00% EER is impressive, but several cells (e.g., replacement at L=0.25 s) show EERs around 33–38%; error bars would help assess the stability of the comparison across the 10 test speakers.
[§2.3, partition definition] The phrase 'Query frames whose HSTW alignment lies entirely in the tampering plane' is ambiguous. Does it mean the query frame is aligned while the HSTW path is in the tampered state, or that the whole path segment is in that plane? Clarify, since the state assignment is central to the proposed fix for the H2 likelihood.

Circularity Check

0 steps flagged · score 2.0 of 10

No material circularity: alignment estimates and the likelihood-ratio score are computed from held-out query observations; self-citations are component citations, not load-bearing.

full rationale

DA-LRT's derivation is self-contained with respect to its classification target. Stage 2 estimates the H1 diagonal alignment and the H2 HSTW alignment directly from the MFCC cost matrix. Stage 3 defines partitions A and B from the difference of these two alignments, estimates the matching-difference Gaussians from partition A, and evaluates the log-likelihood ratio in Eq. 2 on partition B. The final tampering score is therefore not a fitted label, not a benchmark-derived parameter, and not equal to either alignment by construction; Eq. 2 is a nontrivial function of the observed feature differences under two distinct alignment coordinates. The held-out split (half of the speakers) provides independent test data, so the reported EERs are not forced by training-set fitting. The self-citations ([27] audio hashprints and [28] HSTW) are used as component algorithms and are re-described inside the paper; no uniqueness theorem from the same authors is invoked to forbid alternatives. The paper itself acknowledges the independence assumptions are faulty ('Clearly, these assumptions are faulty...'), which flags a real validity limitation: the H2 non-matching model is estimated from matching partition A and may be miscalibrated for actual insertion/deletion/replacement statistics. That is the skeptic's Eq. 1-2 concern, but it is a modeling weakness rather than a circular reduction: it does not make the predicted tampering score identical to its inputs. Hence no material circularity; the score of 2 reflects only the minor presence of not-load-bearing self-citations.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central result rests on a handful of alignment penalties and statistical assumptions; most are hand-set or estimated from the same data used for scoring, so the method's success is not fully determined by an externally measured parameter set.

free parameters (5)

HSTW insertion penalty α
Dynamic programming penalty for insertion transitions; value never reported in §2.2, and no optimization procedure is described.
HSTW deletion penalty γ
Dynamic programming penalty for deletion transitions; value never reported in §2.2, and no optimization procedure is described.
HSTW plane-transition penalty β
Penalty for switching between matching and tampered states; value never reported in §2.2.
Frame-agreement threshold = ≥2 frames
Query frames are placed in partition B when alignments differ by 2 or more frames; threshold chosen by hand in §2.3.
Per-query match statistics μ_f, σ_f², σ̃_f² = estimated per query from partition A
Gaussian parameters for matching and non-matching feature differences are fit to the same query's agreeing frames; central to the LRT.

assumptions (6)

domain assumption A non-tampered query aligns to the reference along a straight diagonal with a single unknown offset.
Used to define H1 alignment in §2.2; if the source recording itself has edits or variable speed, this fails.
ad hoc to paper The two-state HSTW model (matching vs tampered, transitions (1,1), (0,1), (1,0)) captures all relevant tampering operations.
Assumed in §2.2; replacement/insertion/deletion are modeled, but other edits (speed change, re-encoding with drift) are not.
ad hoc to paper In tampered regions, query and reference features are independent, with marginals from matching frames.
Stated in §2.3 to derive H2 distributions; no evidence that tampered regions follow the same marginals.
domain assumption MFCC features with deltas are a sufficient representation of audio content for cross verification.
Used throughout for the cost matrix and hashprint pre-selection; robust to codec/volume but not to all transformations.
ad hoc to paper Features are independent across time and across the 39 MFCC dimensions for the LRT.
Explicitly acknowledged as faulty in §2.3; used for tractability and assumed symmetric.
domain assumption A matching trusted reference recording has already been identified.
Pre-selection assumes this in §2.1 footnote 3; if not, the method cannot verify.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Audio Cross Verification Using Dual Alignment Likelihood Ratio Test." pith.science (2026). https://pith.science/paper/KODR5ZV4

@misc{pith2026260718190,
  author       = {Pith},
  title        = {Pith review of: Audio Cross Verification Using Dual Alignment Likelihood Ratio Test},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KODR5ZV4}},
  note         = {Machine review of arXiv:2607.18190}
}

read the original abstract

This paper explores a way to verify that audio has not been maliciously tampered in a specific context: short viral videos taken from news recordings. Rather than trying to detect artifacts of tampering (internal inconsistency), we focus on positively verifying a query against a trusted source such as a news recording (external consistency). We propose a method for cross verifying a short audio query against a reference recording from which it was taken. Our approach is to define two hypotheses (non-tampered vs tampered), calculate the most likely alignment between query and reference for each hypothesis, and then perform a likelihood ratio test on the two alignments. We show that this method is fast to compute, much more robust than using MFCC features with Euclidean distance, and has the key benefit of explainability. Our cross verification approach provides an alternative perspective and complementary tool to existing tampering detection methods.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 3 linked inside Pith

[1]

Verified against NBC

INTRODUCTION One significant issue in our society today is the reliability of audio- visual information. The availability of deep fake technology and au- dio/video digital editing software has made it easy for non-experts to generate or modify audiovisual content in a way that seems realistic. These technologies have been used for nefarious purposes such ...

arXiv 2023
[2]

No match found with NBC

SYSTEM DESCRIPTION The DA-LRT method has three main stages (see Figure 1), which are described in detail in the next three subsections.2 Our goal is to com- pare a short query recording to a reference recording (from which it was taken), and to determine if the query matches the reference or if it has been tampered. 2.1. Pre-Selection The first stage is p...
[3]

matching

(The MFCC features are stored for later use in stages 2 and 3.) Third, we identify the offset in the reference recording that results in the lowest total Hamming distance between the corresponding bi- nary feature sequences (assuming a 1-to-1 correspondence). This can be computed efficiently by encoding each frame in memory as a single 32-bit integer and ...
[4]

This dataset contains high-quality audio recordings of 20 different speakers each reading 5 scripts, where scripts are 2–4 minutes long

EXPERIMENTAL SETUP We used the DAPS dataset [29] for our experiments. This dataset contains high-quality audio recordings of 20 different speakers each reading 5 scripts, where scripts are 2–4 minutes long. The dataset also includes recordings in multiple acoustic conditions, which will allow us to study the effect of acoustic environment in future work. ...

2000
[5]

MFCC-Euclidean

RESULTS Table 1 shows the results of our proposed method on the audio cross verification task. The left half of Table 1 shows the performance of the proposed system, and the right half shows the performance of a baseline system (“MFCC-Euclidean”) that uses the Euclidean distance between MFCC features as a tampering score. The table shows the EER of both s...
[6]

ACKNOWLEDGMENTS This material is based upon work supported by the National Science Foundation under Grant No. 1948531
[7]

The emergence of deepfake technology: A review,

Mika Westerlund, “The emergence of deepfake technology: A review,”Technology Innovation Management Review, vol. 9, no. 11, 2019

2019
[8]

Deepfakes and dis- information: Exploring the impact of synthetic political video on deception, uncertainty, and trust in news,

Cristian Vaccari and Andrew Chadwick, “Deepfakes and dis- information: Exploring the impact of synthetic political video on deception, uncertainty, and trust in news,”Social Media+ Society, vol. 6, no. 1, pp. 2056305120903408, 2020

2020

Show all 35 references

[9]

Angular margin softmax loss and its variants for double compressed amr audio detection,

Aykut B ¨uker and Cemal Hanilc ¸i, “Angular margin softmax loss and its variants for double compressed amr audio detection,” in Proceedings of the 2021 ACM Workshop on Information Hid- ing and Multimedia Security, 2021, pp. 45–50

2021
[10]

Detection of dou- ble compressed amr audio using stacked autoencoder,

Da Luo, Rui Yang, Bin Li, and Jiwu Huang, “Detection of dou- ble compressed amr audio using stacked autoencoder,”IEEE Transactions on Information Forensics and Security, vol. 12, no. 2, pp. 432–444, 2016

2016
[11]

Tic-tac, forgery time has run-up! live acoustic watermarking for integrity check in forensic applications,

Valentin A Nita and Amelia Ciobanu, “Tic-tac, forgery time has run-up! live acoustic watermarking for integrity check in forensic applications,” in2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, pp. 1977–1981

2018
[12]

TIC-TAC based live acoustic watermarking with improved forgery detection per- formances,

RA Dobre, RO Preda, and AE Marcu, “TIC-TAC based live acoustic watermarking with improved forgery detection per- formances,” inIEEE 25th International Symposium for Design and Technology in Electronic Packaging (SIITME), 2019, pp. 408–412

2019
[13]

ENF based video forgery detection algorithm,

Yufei Wang, Yongjian Hu, Alan Wee-Chung Liew, and Chang- Tsun Li, “ENF based video forgery detection algorithm,”In- ternational Journal of Digital Crime and Forensics (IJDCF), vol. 12, no. 1, pp. 131–156, 2020

2020
[14]

Supervised audio tampering detection using an autoregressive model,

Xiaodan Lin and Xiangui Kang, “Supervised audio tampering detection using an autoregressive model,” in2017 IEEE inter- national conference on acoustics, speech and signal processing (ICASSP), 2017, pp. 2142–2146

2017
[15]

Tampering detection in digital audio recording based on statistical reverberation fea- tures,

Tejas Bhangale and Rashmika Patole, “Tampering detection in digital audio recording based on statistical reverberation fea- tures,” inSoft Computing and Signal Processing, pp. 583–591. 2019

2019
[16]

Detecting audio splic- ing forgery algorithm based on local noise level estimation,

Xuebo Meng, Chen Li, and Lihua Tian, “Detecting audio splic- ing forgery algorithm based on local noise level estimation,” in 2018 5th international conference on systems and informatics (ICSAI), 2018, pp. 861–865

2018
[17]

A robust deep audio splicing detection method via singularity detection fea- ture,

Kanghao Zhang, Shan Liang, Shuai Nie, Shulin He, Jiahui Pan, Xueliang Zhang, Haoxin Ma, and Jiangyan Yi, “A robust deep audio splicing detection method via singularity detection fea- ture,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 20...

2022
[18]

A deep learning approach for splicing detection in digital audios,

Akanksha Chuchra, Mandeep Kaur, and Savita Gupta, “A deep learning approach for splicing detection in digital audios,” in Congress on Intelligent Systems, 2022, pp. 543–558

2022
[19]

LightCvT: Audio forgery detection via fusion of light cnn and transformer,

Chenyu Liu, Jia Li, Junxian Duan, Haifeng Shen, and Huaibo Huang, “LightCvT: Audio forgery detection via fusion of light cnn and transformer,” in2021 10th International Conference on Computing and Pattern Recognition, 2021, pp. 99–105

2021
[20]

Audio forgery detection techniques: present and past review,

Prabhu R Bevinamarad and MS Shirldonkar, “Audio forgery detection techniques: present and past review,” in2020 4th International Conference on Trends in Electronics and Infor- matics (ICOEI), 2020, pp. 613–618

2020
[21]

The partialspoof database and coun- termeasures for the detection of short generated audio seg- ments embedded in a speech utterance,

Lin Zhang, Xin Wang, Erica Cooper, Nicholas Evans, and Junichi Yamagishi, “The partialspoof database and coun- termeasures for the detection of short generated audio seg- ments embedded in a speech utterance,”arXiv preprint arXiv:2204.05177, 2022

2022 arXiv
[22]

Do you really mean that? content driven audio-visual deepfake dataset and multimodal method for temporal forgery localization,

Zhixi Cai, Kalin Stefanov, Abhinav Dhall, and Munawar Hayat, “Do you really mean that? content driven audio-visual deepfake dataset and multimodal method for temporal forgery localization,”arXiv preprint arXiv:2204.06228, 2022

2022 arXiv
[23]

Anti- forensics of fake stereo audio using generative adversarial net- work,

Tianyun Liu, Diqun Yan, Nan Yan, and Gang Chen, “Anti- forensics of fake stereo audio using generative adversarial net- work,”Multimedia Tools and Applications, vol. 81, no. 12, pp. 17155–17167, 2022

2022
[24]

Anti- forensics of double compressed mp3 audio,

Biaoli Tao, Rangding Wang, Diqun Yan, and Chao Jin, “Anti- forensics of double compressed mp3 audio,”International Journal of Digital Crime and Forensics (IJDCF), vol. 12, no. 3, pp. 45–57, 2020

2020
[25]

Detection of speech smoothing on very short clips,

Qi Yan, Rui Yang, and Jiwu Huang, “Detection of speech smoothing on very short clips,”IEEE Transactions on Infor- mation Forensics and Security, vol. 14, no. 9, pp. 2441–2453, 2019

2019
[26]

Anti- forensics of audio source identification using generative ad- versarial network,

Xiaowen Li, Diqun Yan, Li Dong, and Rangding Wang, “Anti- forensics of audio source identification using generative ad- versarial network,”IEEE Access, vol. 7, pp. 184332–184339, 2019

2019
[27]

Error analysis of forensic enf matching,

Guang Hua, “Error analysis of forensic enf matching,” inIEEE International Workshop on Information Forensics and Security (WIFS), 2018, pp. 1–7

2018
[28]

Audio authentication by exploring the absolute-error- map of enf signals,

Guang Hua, Ying Zhang, Jonathan Goh, and Vrizlynn LL Thing, “Audio authentication by exploring the absolute-error- map of enf signals,”IEEE Transactions on Information Foren- sics and Security, vol. 11, no. 5, pp. 1003–1016, 2016

2016
[29]

Audio forensic authentication based on mocc between enf and reference signals,

Zhisheng Lv, Yongjian Hu, Chang-Tsun Li, and Bei-bei Liu, “Audio forensic authentication based on mocc between enf and reference signals,” inIEEE China Summit and International Conference on Signal and Information Processing, 2013, pp. 427–431

2013
[30]

Applications of enf criterion in forensic au- dio, video, computer and telecommunication analysis,

Catalin Grigoras, “Applications of enf criterion in forensic au- dio, video, computer and telecommunication analysis,”Foren- sic science international, vol. 167, no. 2-3, pp. 136–145, 2007

2007
[31]

An industrial strength audio search algorithm,

Avery Wang, “An industrial strength audio search algorithm,” inProceedings of the International Society for Music Informa- tion Retrieval Conference (ISMIR), 2003, pp. 7–13

2003
[32]

Audio fingerprinting: Combining computer vision & data stream processing,

Shumeet Baluja and Michele Covell, “Audio fingerprinting: Combining computer vision & data stream processing,” in IEEE International Conference on Acoustics, Speech and Sig- nal Processing (ICASSP), 2007, vol. 2, pp. 213–216

2007
[33]

Known-artist live song identification using audio hashprints,

TJ Tsai, Thomas Pr ¨atzlich, and Meinard M¨uller, “Known-artist live song identification using audio hashprints,”IEEE Trans- actions on Multimedia, vol. 19, no. 7, pp. 1569–1582, 2017

2017
[34]

Parameter-free ordered partial match alignment with hidden state time warping,

Claire Chang, Thaxter Shaw, Arya Goutam, Christina Lau, Mengyi Shan, and TJ Tsai, “Parameter-free ordered partial match alignment with hidden state time warping,”Applied Sci- ences, vol. 12, no. 8, pp. 3783, 2022

2022
[35]

Gautham J Mysore, “Can we automatically transform speech recorded on common consumer devices in real-world environ- ments into professional production quality speech?—a dataset, insights, and challenges,”IEEE Signal Processing Letters, vol. 22, no. 8, pp. 1006–1010, 2014

2014

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Verified against NBC

INTRODUCTION One significant issue in our society today is the reliability of audio- visual information. The availability of deep fake technology and au- dio/video digital editing software has made it easy for non-experts to generate or modify audiovisual content in a way that seems realistic. These technologies have been used for nefarious purposes such ...

arXiv 2023

[2] [2]

No match found with NBC

SYSTEM DESCRIPTION The DA-LRT method has three main stages (see Figure 1), which are described in detail in the next three subsections.2 Our goal is to com- pare a short query recording to a reference recording (from which it was taken), and to determine if the query matches the reference or if it has been tampered. 2.1. Pre-Selection The first stage is p...

[3] [3]

matching

(The MFCC features are stored for later use in stages 2 and 3.) Third, we identify the offset in the reference recording that results in the lowest total Hamming distance between the corresponding bi- nary feature sequences (assuming a 1-to-1 correspondence). This can be computed efficiently by encoding each frame in memory as a single 32-bit integer and ...

[4] [4]

This dataset contains high-quality audio recordings of 20 different speakers each reading 5 scripts, where scripts are 2–4 minutes long

EXPERIMENTAL SETUP We used the DAPS dataset [29] for our experiments. This dataset contains high-quality audio recordings of 20 different speakers each reading 5 scripts, where scripts are 2–4 minutes long. The dataset also includes recordings in multiple acoustic conditions, which will allow us to study the effect of acoustic environment in future work. ...

2000

[5] [5]

MFCC-Euclidean

RESULTS Table 1 shows the results of our proposed method on the audio cross verification task. The left half of Table 1 shows the performance of the proposed system, and the right half shows the performance of a baseline system (“MFCC-Euclidean”) that uses the Euclidean distance between MFCC features as a tampering score. The table shows the EER of both s...

[6] [6]

ACKNOWLEDGMENTS This material is based upon work supported by the National Science Foundation under Grant No. 1948531

[7] [7]

The emergence of deepfake technology: A review,

Mika Westerlund, “The emergence of deepfake technology: A review,”Technology Innovation Management Review, vol. 9, no. 11, 2019

2019

[8] [8]

Deepfakes and dis- information: Exploring the impact of synthetic political video on deception, uncertainty, and trust in news,

Cristian Vaccari and Andrew Chadwick, “Deepfakes and dis- information: Exploring the impact of synthetic political video on deception, uncertainty, and trust in news,”Social Media+ Society, vol. 6, no. 1, pp. 2056305120903408, 2020

2020