Pith. sign in

REVIEW 4 major objections 5 minor 42 references

Double Entendre: Robust Audio-Based AI-Generated Lyrics Detection via Multi-View Fusion

T0 review · 4 major / 5 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 AI-generated lyrics can be detected from audio alone with 94.9% recall.

desk verdict Solid applied paper: first audio-only pipeline combining ASR lyrics and speech embeddings for AI-lyrics detection, but the partly-fake control leaves a renderer confound unresolved; worth refereeing. read the letter →

arxiv 2506.15981 v2 pith:E5AGGGNE submitted 2025-06-19 cs.CL cs.AIcs.SDeess.AS

Markus Frohmann , Gabriel Meseguer-Brocal , Markus Schedl , Elena V. Epure This is my paper
classification cs.CLcs.AIcs.SDeess.AS
keywords AI-generatedmusicdetectionlyricsauthorshipmultimodallatefusionspeechembeddingsautomatictranscriptionaudioperturbationrobustnessout-of-domaingeneralization
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

AI music generators such as Suno and Udio can now create complete songs, but existing detectors either need clean lyric text (often unavailable) or fail under audio perturbations and unseen generators. This paper proposes DE-detect, a detector that takes only the audio of a song and combines two views: the lyrics transcribed by Whisper and embedded with LLM2Vec, and speech embeddings from XEUS that capture how the words are sung. On the authors' multilingual dataset, DE-detect reaches 94.9% macro-recall and 98.5% AUROC in-domain, beating every lyrics-based baseline including one given ground-truth transcripts, and holds 94.1% recall on an unseen generator (Udio) and across five audio attacks. The practical consequence is that platforms can flag AI-written lyrics on real uploaded tracks without relying on lyric metadata or fragile audio artifacts.

What carries the argument

The load-bearing mechanism is a late fusion of two frozen feature extractors: Whisper large-v2 transcribes the audio to lyrics, which LLM2Vec (Llama3 8B) embeds as a single text vector (the 'what'), while XEUS speech embeddings are mean-pooled into a single audio vector (the 'how', capturing prosody, intonation, and timbre). Each vector is linearly projected to 128 dimensions, concatenated, and classified by a small MLP. The fusion is what gives the system its edge: text alone reaches 90.7% recall, speech alone 92.2%, and the combination 94.9%, and the fusion is what keeps the method stable when the audio is perturbed or comes from an unseen generator.

What would settle it

Compute a perceptual or spectral distance between Suno audio that sings human lyrics and Suno audio that sings AI lyrics, matched for genre and language; if a simple acoustic classifier separates the two above chance, then the 'mostly similar artifacts' premise fails and the model's speech branch may be exploiting rendering differences rather than lyrical authorship. A second check would be to run the partly-fake experiment on Udio once it supports lyric conditioning, since out-of-domain rendering should not preserve the same artifact signature.

Watch

Extended reading notes

Core claim

The paper's central claim is that AI-generated lyrics can be detected from audio alone, without clean lyric text, by combining two complementary views of the same recording: the semantic content of the transcribed lyrics and the acoustic-linguistic information in the singing voice. On the authors' dataset of 7,190 songs, the resulting system, DE-detect, achieves 94.9% macro-recall and 98.5% AUROC in-domain, outperforming all lyrics-based baselines including one using ground-truth non-transcribed lyrics, and maintains 94.1% recall on an out-of-domain generator (Udio) and under five audio perturbations. The authors argue that the speech branch is not merely detecting synthetic audio artifacts, because it performs at chance when separating real human recordings from Suno-rendered songs with human lyrics, while distinguishing AI-lyric from human-lyric songs rendered by the same synthesizer.

Load-bearing premise

The partly-fake control assumes audio artifacts from Suno's synthesizer are similar whether the sung lyrics are human-written or AI-written; if Suno's rendering of AI-written lyrics carries a consistent acoustic signature, the model might be detecting the rendering rather than the lyric authorship.

Editorial extensions

If this is right

  • Music streaming platforms can flag AI-generated lyrics on newly ingested tracks using only the uploaded audio, without waiting for lyric metadata or clean transcripts.
  • Spectrogram-based audio detectors, which excel in-domain but collapse under pitch change and on unseen generators, are not the only practical option; a lyrics-plus-speech fusion holds up under those same conditions.
  • Because the best text branch (LLM2Vec on Whisper transcripts) nearly matches the clean-lyrics baseline (90.7% vs 94.3% recall), automatic transcription preserves most of the authorship signal, making text-based detection usable in the wild.
  • The speech branch's chance-level performance on real-vs-synthetic audio with human lyrics indicates the model is reading lyrical authorship rather than synthesizer artifacts, which is what would make the approach transfer to future generators.

Reading between the lines

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

  • If the partly-fake control holds, the same late-fusion design could be pointed at other authorship questions where the 'how' complements the 'what,' such as distinguishing ghost-written from self-written lyrics, or detecting machine-translated song lyrics.
  • Because the method works from audio alone, it could be embedded in broadcast or live-stream monitoring, where lyric text is never available, to spot AI-generated songs at ingestion time.
  • The cleanest stress test the paper did not run is a combined attack (e.g., pitch shift plus time stretch); the authors note this gap, and such combined perturbations may be the natural adversarial next step for anyone evaluating the method.
  • A future generator that produces vocals with a different rendering style (or an artist re-singing AI lyrics) would reveal how much of the speech branch's signal is tied to Suno's text-conditioned synthesis.
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

4 major / 5 minor

Summary. The paper proposes DE-detect, a late-fusion pipeline that combines automatically transcribed lyrics (Whisper large-v2 + LLM2Vec text embeddings) with speech embeddings (XEUS) to detect AI-generated lyrics from audio alone. The authors construct a dataset by pairing human and LLM-generated lyrics from Labrak et al. (2025) with Suno v3.5 audio, and evaluate in-domain, under five audio perturbations, and on out-of-domain Udio audio. They report that DE-detect reaches 94.9% macro-recall and 98.5% AUROC in-domain, outperforms lyrics-only baselines (including a ground-truth-lyrics LLM2Vec baseline), and remains robust to perturbations and to the Udio domain. The paper includes a 'partly-fake' control condition where Suno audio is generated from human lyrics, intended to separate lyrical-content detection from audio-artifact detection. The key claims are that the method detects AI-written lyrics rather than generic AI audio artifacts, that the multi-view fusion improves over unimodal components, and that the pipeline is practically applicable since it needs only raw audio.

Significance. If the central claims hold, the paper addresses a real practical gap: existing lyrics-based detectors assume clean text, while audio artifact detectors overfit to specific generators. The proposed audio-only, multimodal late-fusion design is modular and the authors provide code and reproducibility details. The work also ships a useful control condition (partly-fake) and an out-of-domain Udio evaluation, which go beyond most AIGM detection studies. The main contributions—demonstrating that fused transcribed-text and speech embeddings improve robustness over unimodal detectors, and that transcribed lyrics retain enough signal to rival ground-truth lyrics—are meaningful for the music information retrieval and AI-generated content detection communities. However, the significance is conditional on resolving a confound in the partly-fake control and on establishing that the reported performance differences are not artifacts of evaluation-set component selection or missing variance estimates.

major comments (4)
  1. [§3.1, Table 2] The partly-fake control is the load-bearing evidence for the paper's central claim that the model detects AI-written lyrics rather than synthetic vocal rendering, but the control does not actually hold the renderer fixed across lyric authorship. The paper assumes in §3.1 that audio artifacts 'should be mostly similar' for fully fake and partly-fake songs because both use Suno audio. However, Suno synthesizes audio conditioned on the lyric text itself, and LLM-written lyrics differ systematically from human lyrics in statistical text properties (repetition, line length, token usage). These properties can change prosody, enunciation, and even non-linguistic vocal artifacts, all of which XEUS—a speech embedding model explicitly described as capturing prosody and intonation—could exploit. Table 2 shows XEUS at 50.5% recall in real-vs-partly-fake (chance, suggesting it is not detecting generic Suno artifacts) but 92.0% in fake-vs-partly-fake; this pattern is consistent with the model detecting renderer-induced acoustic correlates of AI-written text, not linguistic authorship per se. The Udio out-of-domain evaluation does not resolve this because Udio also conditions on lyrics. A control that holds the vocal renderer fixed across human and AI lyrics (e.g., same singer/synthesis voice reading both text types with minimal prosodic variation, or a text-to-speech renderer applied to both) would be needed to separate the two hypotheses.
  2. [§4.1, Fig. 3, Table 1, Table 3] The unimodal components are selected based on recall scores computed on what appears to be the same evaluation split used for the final reported results, introducing selection bias. Section 4.1 states that the authors 'evaluate several unimodal features to select as components' and then choose Whisper large-v2, LLM2Vec, and XEUS based on Figures 2–3 and Table 1, using the same in-domain data that later produces the headline numbers in Table 3. This is a form of peeking at the test set and inflates the apparent advantage of the selected fusion. The paper does not report a fixed validation split or nested evaluation, and it reports no error bars or multiple-seed variance anywhere; the claimed gains of DE-detect over XEUS alone (~2.7 recall points) and over LLM2Vec-transcript (~4.2 points) are small relative to what selection on the evaluation set could explain. At minimum, the authors should either use a held-out validation set for component selection, report results across multiple seeds with mean and standard deviation, or show that the selection decision is insensitive to the evaluation split.
  3. [§4.2, Table 3] The paper does not provide statistical significance tests or confidence intervals for the key differences it relies on. The central quantitative claim is that DE-detect 'outperforms existing lyrics-based detectors' and 'is more robust to audio perturbations.' Table 3 shows DE-detect at 94.9% recall versus LLM2Vec-transcript at 90.7% and XEUS at 92.2%, and Table 4 shows gains of 1.5–2% over the unimodal speech branch under perturbations. With only a single run and no error bars, the reader cannot assess whether these margins are stable. Given the paper's emphasis on modularity and real-world robustness, reporting variance across at least a few random seeds and a paired test (e.g., McNemar or bootstrap over songs) is necessary to support the comparative claims.
  4. [§2, 'Speech Branch'] The paper's motivation for the speech branch is that it captures 'lyrics-related information present only in audio,' but the mechanism is never pinned down empirically. The authors attribute XEUS's strength to singing-voice training data and prosody/timbre capture, yet the only direct evidence for what the speech branch encodes is the partly-fake comparison in Table 2, which, as discussed above, is confounded. The paper would be strengthened by an analysis of which dimensions of XEUS embeddings change between fake and partly-fake (e.g., probing for prosodic features or comparing embeddings of identical text sung with different renderers). Without such analysis, the claim that the speech branch contributes 'lyrics-related' rather than 'renderer-related' information is not established.
minor comments (5)
  1. [Appendix B] In the description of UAR models, 'trained on texts from 1 million and 5 Reddit users' is likely a typo for '5 million Reddit users'; please clarify.
  2. [Throughout] The model name is inconsistently typeset: 'DE-DETECT', 'DE- DETECT', and 'DE- DETECT' appear. Please normalize the formatting.
  3. [Tables 3 and 4] The header 'UDIO' in Table 4 is a typo; it should read 'UDIO' or 'Udio' consistently.
  4. [§3.1] The sentence describing the partly-fake subset notes that 'a few songs were blocked during generation, making ourPartly-Fake subset slightly smaller.' This mismatch is not quantified; please state the exact number of partly-fake songs used and confirm the class balance in the evaluation split.
  5. [Appendix A] The paper says 'To make sure no encoding-specific patterns are picked up, we convert all audio to mp3 with 128kbps.' It would be helpful to state whether this conversion is also applied to the human recordings and to the Udio test audio, so that the reader can assess whether encoding is a controlled variable across conditions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported gains are empirical, benchmarked against external baselines, Udio transfer, and audio perturbations; self-citation is data provenance, not a load-bearing derivation.

full rationale

This paper is an empirical evaluation, not a derivation from first principles. The central claims—94.9% macro-recall and 98.5% AUROC—are produced by training an MLP on audio-derived text and speech features and testing on held-out songs, with additional out-of-domain Udio data and perturbed audio as external checks. The partly-fake control is an experimental design intended to separate lyrical content from audio artifacts; while it relies on an untested assumption that Suno artifacts are 'mostly similar' for partly-fake and fully fake songs, that is a validity limitation or confound, not a circular reduction: the conclusion is not identical to the assumption by construction. The use of the Labrak et al. (2025) dataset, on which three of the present authors are co-authors, is a data-provenance overlap rather than a load-bearing theorem; the paper invokes no uniqueness result and no fitted parameter is renamed as a prediction. Component selection (Whisper large-v2, LLM2Vec, XEUS) was performed on the evaluation set, which is a mild self-referential optimization, but it does not make the reported performance equivalent to the selection criterion and is not a circular step. No equation reduces to its inputs by definition, so the appropriate finding is no significant circularity.

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

The central empirical claim rests on three domain assumptions: ASR preserves lyrical authorship signals, speech embeddings capture generative singing cues, and the Suno-based dataset isolates lyric authorship from audio generation. The partly-fake experiment is the main check on the last point, but it cannot fully remove the distribution gap between original recordings and Suno-rendered audio.

free parameters (3)
  • Unimodal component selection (transcriber, text encoder, speech encoder) = Whisper large-v2, LLM2Vec Llama3 8B, XEUS
    Chosen after comparing recall scores in Section 4.1; no separate validation split is described, so selection appears to use the evaluation set and may inflate final results.
  • Fusion projection dimension = 128
    Used to down-project each branch before concatenation (Appendix A); no ablation of this dimension is reported.
  • MLP hidden sizes = 256, 128 for unimodal; 128 after fusion
    Architecture choices from Appendix A; authors state variant choices made no significant difference.
assumptions (4)
  • domain assumption Transcription with Whisper large-v2 preserves AI-visible lyrical characteristics well enough for detection.
    Section 4.2 states 'transcription effectively retains AI-generated lyric characteristics for detection'; if ASR errors erase or add stylistic markers, the text branch would not measure lyrics authorship.
  • domain assumption Speech embeddings from XEUS capture paralinguistic cues (prosody, intonation, timbre) indicative of AI-generated sung lyrics, and not merely low-level audio artifacts.
    Section 2 hypothesizes 'speech embeddings capture this how'; Table 2 supports via chance-level real vs partly-fake performance, but this is an empirical assumption about representation quality.
  • ad hoc to paper Audio artifacts in Suno-generated audio are mostly similar for AI and human lyrics, so the partly-fake condition isolates lyrical content.
    Section 3.1: 'This mitigates the influence of audio artifacts that should be mostly similar for partly-fake and fully fake'; this is a domain assumption specific to the experimental design, not a proven fact.
  • domain assumption The original human recordings and the Suno-generated fake recordings differ predominantly in lyrics authorship rather than in recording conditions, vocal quality, or genre distribution.
    The dataset pairs real songs with original audio and AI-lyrics songs with Suno audio (Section 3.1); this confound is acknowledged in Limitations as 'Suno-centric' bias.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Double Entendre: Robust Audio-Based AI-Generated Lyrics Detection via Multi-View Fusion." pith.science (2026). https://pith.science/paper/E5AGGGNE

@misc{pith2026250615981,
  author       = {Pith},
  title        = {Pith review of: Double Entendre: Robust Audio-Based AI-Generated Lyrics Detection via Multi-View Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E5AGGGNE}},
  note         = {Machine review of arXiv:2506.15981}
}
read the original abstract

The rapid advancement of AI-based music generation tools is revolutionizing the music industry but also posing challenges to artists, copyright holders, and providers alike. This necessitates reliable methods for detecting such AI-generated content. However, existing detectors, relying on either audio or lyrics, face key practical limitations: audio-based detectors fail to generalize to new or unseen generators and are vulnerable to audio perturbations; lyrics-based methods require cleanly formatted and accurate lyrics, unavailable in practice. To overcome these limitations, we propose a novel, practically grounded approach: a multimodal, modular late-fusion pipeline that combines automatically transcribed sung lyrics and speech features capturing lyrics-related information within the audio. By relying on lyrical aspects directly from audio, our method enhances robustness, mitigates susceptibility to low-level artifacts, and enables practical applicability. Experiments show that our method, DE-detect, outperforms existing lyrics-based detectors while also being more robust to audio perturbations. Thus, it offers an effective, robust solution for detecting AI-generated music in real-world scenarios. Our code is available at https://github.com/deezer/robust-AI-lyrics-detection.

Figures

Figures reproduced from arXiv: 2506.15981 by the authors.

Figure 1
Figure 1. Overview of our pipeline to robustly detect AI-generated lyrics when only audio is available. In the top [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Word error rates (WER) of different transcrip [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 15 canonical work pages

  1. [1]

    Darius Afchar, Gabriel Meseguer-Brocal, and Romain Hennequin. 2024. https://api.semanticscholar.org/CorpusID:269614314 Detecting music deepfakes is easy but actually hard . ArXiv, abs/2405.04181

    arXiv 2024

  2. [2]

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. https://proceedings.neurips.cc/paper_files/paper/2020/file/92d1e1eb1cd6f9fba3227870bb6d7f07-Paper.pdf wav2vec 2.0: A framework for self-supervised learning of speech representations . In Advances in Neural Information Processing Systems, volume 33, pages 12449--12460. Curran Associates, Inc

    work page 2020

  3. [3]

    Solon Barocas, Kate Crawford, Aaron Shapiro, and Hanna Wallach. 2017. The problem with bias: Allocative versus representational harms in machine learning. In 9th Annual Conference of the Special Interest Group for Computing, Information and Society

    work page 2017

  4. [4]

    Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. https://openreview.net/forum?id=IW1PR7vEBf LLM2V ec: Large language models are secretly powerful text encoders . In First Conference on Language Modeling

    2024

  5. [5]

    Daria Beresneva. 2016. Computer-generated text detection using machine learning: A systematic review. In Natural Language Processing and Information Systems, pages 421--426, Cham. Springer International Publishing

    work page 2016

  6. [6]

    Jianlyu Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024 a . https://doi.org/10.18653/v1/2024.findings-acl.137 M 3-embedding: Multi-linguality, multi-functionality, multi-granularity text embeddings through self-knowledge distillation . In Findings of the Association for Computational Linguistics: ACL 2024, pages 2318--2335, Bangk...

    doi:10.18653/v1/2024.findings-acl.137 2024

  7. [7]

    William Chen, Wangyou Zhang, Yifan Peng, Xinjian Li, Jinchuan Tian, Jiatong Shi, Xuankai Chang, Soumi Maiti, Karen Livescu, and Shinji Watanabe. 2024 b . https://doi.org/10.18653/v1/2024.emnlp-main.570 Towards robust speech representation learning for thousands of languages . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language P...

    work page doi:10.18653/v1/2024.emnlp-main.570 2024

  8. [8]

    Ondrej C \'i fka, Hendrik Schreiber, Luke Miner, and Fabian-Robert St \"o ter. 2024. https://api.semanticscholar.org/CorpusID:271860246 Lyrics transcription for humans: A readability-aware benchmark . ArXiv, abs/2408.06370

    arXiv 2024

Show all 42 references
  1. [9]

    Kao, Ann Lee, Xutai Ma, Alexandre Mourachko, Benjamin Peloquin, Juan Pino, Sravya Popuri, Christophe Ropers, Safiyyah Saleem, Holger Schwenk, Anna Y

    Seamless Communication, Lo \"i c Barrault, Yu-An Chung, Mariano Coria Meglioli, David Dale, Ning Dong, Mark Duppenthaler, Paul-Ambroise Duquenne, Brian Ellis, Hady ElSahar, Justin Haaheim, John Hoffman, Min-Jae Hwang, Hirofumi Inaguma, Christopher Klaiber, Ilia Kulikov, Pengwe...

    2023 arXiv

  2. [10]

    Di Cooke, Abigail Edwards, Sophia Barkoff, and Kathryn Kelly. 2024. https://api.semanticscholar.org/CorpusID:268681720 As good as a coin toss: Human detection of ai-generated images, videos, audio, and audiovisual stimuli . ArXiv, abs/2403.16760

    2024 arXiv

  3. [11]

    William Falcon and The PyTorch Lightning team . 2019. https://doi.org/10.5281/zenodo.3828935 PyTorch Lightning

    2019 doi

  4. [12]

    Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.552 S im CSE : Simple contrastive learning of sentence embeddings . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6894--6910, Online ...

    2021 doi

  5. [13]

    Sebastian Gehrmann, Hendrik Strobelt, and Alexander Rush. 2019. https://doi.org/10.18653/v1/P19-3019 GLTR : Statistical detection and visualization of generated text . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrat...

    2019 doi

  6. [14]

    Antoine Henry, Valdy Wiratama, Adelaida Afilipoaie, Heritiana Ranaivoson, and Eric Arrivé. 2024. https://doi.org/10.1177/27523543241269047 Impacts of ai on music consumption and fairness . Emerging Media, 2(3):382--396

    2024 doi

  7. [15]

    Daphne Ippolito, Daniel Duckworth, Chris Callison-Burch, and Douglas Eck. 2020. https://doi.org/10.18653/v1/2020.acl-main.164 Automatic detection of generated text is easiest when humans are fooled . In Proceedings of the 58th Annual Meeting of the Association for Computationa...

    2020 doi

  8. [16]

    Guillaume Klein, Jong Wook Kim, Yoon Kim, and Clement Delangue. 2023. https://doi.org/10.5281/zenodo.10084808 faster-whisper: A reimplementation of openai's whisper model using ctranslate2

    2023 doi

  9. [17]

    Yanis Labrak, Markus Frohmann, Gabriel Meseguer-Brocal, and Elena V. Epure. 2025. https://aclanthology.org/2025.trustnlp-main.34/ Synthetic lyrics detection across languages and genres . In Proceedings of the 5th Workshop on Trustworthy NLP (TrustNLP 2025), pages 524--541, Alb...

    2025

  10. [18]

    Thomas Lavergne, Tanguy Urvoy, and Fran c ois Yvon. 2008. Detecting fake content with relative entropy scoring. In Proceedings of the 2008 International Conference on Uncovering Plagiarism, Authorship and Social Software Misuse - Volume 377, PAN'08, page 27–31, Aachen, DEU. CE...

    2008

  11. [19]

    Yafu Li, Qintong Li, Leyang Cui, Wei Bi, Zhilin Wang, Longyue Wang, Linyi Yang, Shuming Shi, and Yue Zhang. 2024 a . https://aclanthology.org/2024.acl-long.3 MAGE : Machine-generated text detection in the wild . In Proceedings of the 62nd Annual Meeting of the Association for ...

    2024

  12. [20]

    Schuller

    Yupei Li, Manuel Milling, Lucia Specia, and Bj \" o rn W. Schuller. 2024 b . https://doi.org/10.48550/ARXIV.2412.00571 From audio deepfake detection to ai-generated music detection - A pathway and overview . CoRR, abs/2412.00571

    2024 doi arXiv

  13. [21]

    Ilya Loshchilov and Frank Hutter. 2017. https://api.semanticscholar.org/CorpusID:53592270 Decoupled weight decay regularization . In International Conference on Learning Representations

    2017

  14. [22]

    Brian McFee, Colin Raffel, Dawen Liang, Daniel P. W. Ellis, Matt McVicar, Eric Battenberg, and Oriol Nieto. 2015. https://api.semanticscholar.org/CorpusID:33504 librosa: Audio and music signal analysis in python . In SciPy

    2015

  15. [23]

    Alessandra Micalizzi. 2024. Artificial Creativity. Perceptions and Prejudices on AI Music Production, pages 481--491. Springer Nature Singapore, Singapore

    2024

  16. [24]

    Manning, and Chelsea Finn

    Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D. Manning, and Chelsea Finn. 2023. Detectgpt: zero-shot machine-generated text detection using probability curvature. In Proceedings of the 40th International Conference on Machine Learning, ICML'23. JMLR.org

    2023

  17. [25]

    Preslav Nakov, Sara Rosenthal, Zornitsa Kozareva, Veselin Stoyanov, Alan Ritter, and Theresa Wilson. 2013. https://aclanthology.org/S13-2052 S em E val-2013 task 2: Sentiment analysis in T witter . In Second Joint Conference on Lexical and Computational Semantics (* SEM ), Vol...

    2013

  18. [26]

    Emilia Parada-Cabaleiro, Maximilian Mayerl, Stefan Brandl, Marcin Skowron, Markus Schedl, Elisabeth Lex, and Eva Zangerle. 2024. https://doi.org/10.1038/s41598-024-55742-x Song lyrics have become simpler and more repetitive over the last five decades . Scientific Reports, 14(1):5531

    2024 doi

  19. [27]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu ...

    2019

  20. [28]

    Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Mamdouh Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, and Michael Auli. 2023. https://api.semanticscholar....

    2023 arXiv

  21. [29]

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. https://api.semanticscholar.org/CorpusID:252923993 Robust speech recognition via large-scale weak supervision . ArXiv, abs/2212.04356

    2022 arXiv

  22. [30]

    Md Awsafur Rahman, Zaber Ibn Abdul Hakim, Najibul Haque Sarker, Bishmoy Paul, and Shaikh Anowarul Fattah. 2024. https://api.semanticscholar.org/CorpusID:271957458 Sonics: Synthetic or not - identifying counterfeit songs . ArXiv, abs/2408.14080

    2024 arXiv

  23. [31]

    Nils Reimers and Iryna Gurevych. 2019. https://doi.org/10.18653/v1/D19-1410 Sentence- BERT : Sentence embeddings using S iamese BERT -networks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference...

    2019 doi

  24. [32]

    Rivera-Soto, Olivia Elizabeth Miano, Juanita Ordonez, Barry Y

    Rafael A. Rivera-Soto, Olivia Elizabeth Miano, Juanita Ordonez, Barry Y. Chen, Aleem Khan, Marcus Bishop, and Nicholas Andrews. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.70 Learning universal authorship representations . In Proceedings of the 2021 Conference on Empiric...

    2021 doi

  25. [33]

    Welty, Christopher A

    Gemma Team Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, L'eonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ram'e, Johan Ferret, Peter Liu, Pouya Dehghani Tafti, Abe Friesen, Michelle Casbon, Sabela Ramos, Ravin Kumar, Charl...

    2024 arXiv

  26. [34]

    C. E. Shannon. 1948. https://doi.org/10.1002/j.1538-7305.1948.tb01338.x A mathematical theory of communication . The Bell System Technical Journal, 27(3):379--423

    1948

  27. [35]

    Weijia Shi, Anirudh Ajith, Mengzhou Xia, Yangsibo Huang, Daogao Liu, Terra Blevins, Danqi Chen, and Luke Zettlemoyer. 2024. https://openreview.net/forum?id=zWqr3MQuNs Detecting pretraining data from large language models . In The Twelfth International Conference on Learning Re...

    2024

  28. [36]

    Peter Sobot. 2021. https://doi.org/10.5281/zenodo.7817838 Pedalboard

    2021 doi

  29. [37]

    Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, Miles McCain, Alex Newhouse, Jason Blazakis, Kris McGuffie, and Jasmine Wang. 2019. https://arxiv.org/abs/1908.09203 Release stra...

    2019 arXiv

  30. [38]

    Wenhui Wang, Hangbo Bao, Shaohan Huang, Li Dong, and Furu Wei. 2021. https://doi.org/10.18653/v1/2021.findings-acl.188 M ini LM v2: Multi-head self-attention relation distillation for compressing pretrained transformers . In Findings of the Association for Computational Lingui...

    2021 doi

  31. [39]

    Shinji Watanabe, Takaaki Hori, Shigeki Karita, Tomoki Hayashi, Jiro Nishitoba, Yuya Unno, Nelson Yalta, Jahn Heymann, Matthew Wiesner, Nanxin Chen, Adithya Renduchintala, and Tsubasa Ochiai. 2018. https://api.semanticscholar.org/CorpusID:4556070 Espnet: End-to-end speech proce...

    2018

  32. [40]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mari...

    2020 doi

  33. [41]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  34. [42]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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