Pith. sign in

REVIEW 4 major objections 5 minor 36 references

When Humans Growl and Birds Speak: High-Fidelity Voice Conversion from Human to Animal and Designed Sounds

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A CVAE-based voice conversion model transforms human speech into animal and designed non-human sounds at 44.1 kHz, beating prior baselines in quality, naturalness, and similarity.

desk verdict Useful engineering extension of non-human VC, but the central MOS claim is contradicted by its own Table 1. read the letter →

arxiv 2505.24336 v1 pith:YUBTNXOF submitted 2025-05-30 eess.AS cs.AIcs.LGcs.SDeess.SP

Minsu Kang , Seolhee Lee , Choonghyeon Lee , Namhyun Cho This is my paper
classification eess.AScs.AIcs.LGcs.SDeess.SP
keywords voiceconversionnon-humananimalsoundssounddesignCVAEhigh-fidelityaudiotimbretransferself-supervisedlearning
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 claims that a conditional variational autoencoder (CVAE) voice conversion model, paired with a 44.1 kHz preprocessing pipeline, can convert ordinary human speech into a wide range of non-human vocalizations—lion roars, birdsong, synthetic growls, and orc- or goblin-style utterances—while preserving the source's meaning and energy. The central bet is that style and linguistic content can be separated well enough to survive an extreme timbre change without needing style labels or animal-specific training data. The authors report that their model beats prior voice conversion systems on most subjective and objective metrics, reaching a similarity MOS of 3.78, an energy-contour correlation of 0.99, and a character error rate of 15.48% on samples with phonetic content. If true, this would let game and film sound designers generate creature voices from a human recording instead of hand-tuning DSP effects.

What carries the argument

The central object is a conditional variational autoencoder (CVAE) voice conversion model whose posterior encoder, decoder, normalizing flow, and prior network are trained jointly with a KL divergence loss, adversarial loss, feature matching loss, and reconstruction loss. Two modifications carry the argument. First, a preprocessing pipeline uses 44.1 kHz audio, a 5 ms STFT hop, and 0-22.05 kHz mel filters to preserve the frequency range and fine temporal detail typical of non-human sounds. Second, the style vector extracted from a reference audio is applied only to the prior network and flow module, not to the posterior encoder or decoder, so the reference timbre does not leak into the acoustic latent. A self-supervised speech model's 12th-layer hidden representation supplies the linguistic encoding, with formant and pitch perturbation expanded to (1.8, 3, 2) to expose the encoder to wide frequency shifts.

What would settle it

Take a set of human screams and growls with no phoneme content, convert them with the proposed model, and ask listeners to match each converted output to its source; if content is not preserved, the conversion claim fails for exactly the non-linguistic sounds the paper highlights.

Watch

Extended reading notes

Core claim

The paper's central claim is that human-to-non-human voice conversion can be extended far beyond dog sounds by redesigning the preprocessing and the CVAE architecture rather than by adding more data or style IDs. Concretely, the authors show that a 5 ms-hop STFT on 44.1 kHz audio with mel filters spanning 0 to 22.05 kHz captures the wide frequency range and transient detail of non-speech voices; feeding the style vector only into the prior network and flow module prevents style overlap with the acoustic encoder; a frequency-domain reconstruction loss recovers abrupt transients; and cosine KL annealing prevents posterior collapse during training. Measured by MOS tests and objective metrics, the converted lion roars, birdsongs, and designed growls are more similar to the target timbre and retain more of the source's linguistic content than outputs from prior baselines.

Load-bearing premise

The model assumes that the 12th-layer hidden representation of a self-supervised speech model trained on human speech still carries the content of screams, growls, and other non-linguistic vocalizations; the paper never checks content preservation on such non-lingual samples.

Editorial extensions

If this is right

  • Game and film sound designers could generate lion roars, birdsong, orc-like growls, and similar creature voices directly from human voice recordings, bypassing manual DSP toolchains.
  • The 44.1 kHz, 5 ms-hop pipeline extends to other non-human sound types beyond the evaluated categories, such as monster screams and zombie utterances, because no style IDs are required during training.
  • Removing the style vector from the acoustic encoder and decoder improves similarity (MOS-S 3.78 vs 3.61), indicating that style overlap is a real bottleneck for extreme timbre conversion.
  • KL annealing is load-bearing: without it, content metrics degrade sharply (CER from 15.48% to 28.89%, WER from 24.02% to 44.69%).

Reading between the lines

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

  • The reliance on human-speech SSL features likely caps performance on pure non-linguistic vocalizations; a content encoder trained on animal and designed sounds, or a contrastive objective on non-speech, is a testable extension that would probe this limit.
  • A weaker baseline scored higher on naturalness (MOS-N 3.22 vs 3.16), suggesting a similarity-naturalness tradeoff: strong timbre conversion may reduce perceived naturalness, and future work could search for a perceptual sweet spot.
  • The same architecture could be reversed—converting animal sounds to human speech or transferring style between two non-human timbres—to test whether the learned latent space is truly timbre-independent.
  • Energy-only prosody, with no f0, may limit conversion of pitched non-human sounds like birdsong; adding a robust pitch estimator for non-harmonic signals is a concrete next step the authors flag as future work.
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. This paper presents H2NH-VC, a conditional-VAE-based voice conversion system that converts human speech into non-human vocalizations including animal sounds (lion roars, birdsong) and designed voices (synthetic growls, orc/goblin-style utterances). The method operates at 44.1 kHz and introduces a preprocessing pipeline with a 5 ms STFT hop, mel-filtering up to 22.05 kHz, expanded information-perturbation ranges, and frame-level energy features. Architectural and training changes include applying the style vector only to the prior and flow modules, using Frequency Domain Reconstruction Loss, and cosine KL annealing. The model is compared with DDDM-VC, Diff-HierVC, and Free-VC on subjective MOS and objective metrics, with ablation studies for the preprocessing, style-embedding placement, and KL annealing.

Significance. If the reported results hold, the work extends voice conversion to a practically relevant and underexplored setting: generating diverse non-human timbres for game audio and sound design. The 44.1 kHz support and the breadth of target sounds go beyond prior dog-only work, and the ablations give internal grounding for the design choices. The demo page is a useful artifact for qualitative assessment. However, the central subjective claim is not currently supported: the naturalness MOS in Table 1 contradicts the abstract, and the evaluation relies on only 9 listeners without significance testing. The objective metrics do not repair this gap because they measure energy correlation and word error on a restricted lingual subset, not perceived naturalness. The contribution is therefore plausible but not yet established at the level claimed.

major comments (4)
  1. [Abstract, Section 4.2.1, Table 1, Conclusion] The abstract and conclusion state that the proposed method outperformed baselines in quality, naturalness, and similarity MOS, but Table 1 shows Free-VC has MOS-N 3.22 while the proposed model scores 3.16. This directly contradicts the central claim. The explanation that Free-VC preserves expressive details because it converts less is a post-hoc interpretation, not a justification for claiming superiority in naturalness. Please correct the claim to reflect the actual direction of each MOS dimension, and add significance tests or confidence intervals so that the remaining reported advantages are statistically supported.
  2. [Section 4.1 'Subjective evaluation'] The MOS evaluation used only 9 listeners and reports no confidence intervals, listener-level variance, or significance tests. Given that the key comparisons involve differences as small as 3.16 vs 3.22 (MOS-N) and 3.16 vs 2.73 (MOS-Q), it is unclear whether any of the claimed improvements are reliable. Please report per-listener statistics and appropriate significance testing (e.g., paired tests with multiple-comparison correction), or clearly state the results as preliminary and without statistical support.
  3. [Section 3.1 'Linguistic Representation' and Section 4.1 'Objective metrics'] The central conversion capability depends on the 12th-layer XLS-R representation preserving content for non-linguistic inputs such as screams, growls, and bird-like vocalizations, but content preservation is measured only with CER/WER on samples that contain lingual phonemic information. This leaves the core task -- converting expressive, non-linguistic vocalizations -- without any content-preservation validation. Please add a content-preservation evaluation on non-linguistic samples (e.g., human yelling, non-verbal screams) or explicitly scope the claim to the subset where content is measurable.
  4. [Section 4.1 'Experimental Settings' and Section 4.2.1] All baselines are trained with conventional speech-focused preprocessing, while the proposed model uses the new preprocessing pipeline. This confounds the effect of the architecture with the effect of the preprocessing in the headline comparison. The w/o PP ablation shows the preprocessing helps the proposed model, but it does not show whether the baselines would benefit equally from the new pipeline. Please train the baselines with the proposed preprocessing as well, or clearly argue why the baseline results would be representative of their best achievable performance under the new pipeline.
minor comments (5)
  1. [Abstract and Figure 1 caption] There are typos in the abstract ('accomodate' should be 'accommodate') and in the Figure 1 caption ('artificailly' should be 'artificially').
  2. [Section 3.2] The paragraph beginning 'As shown in Fig. 2, the proposed model consists...' and the following paragraph starting 'As illustrated in Fig. 2, the proposed model comprises...' describe essentially the same architecture with overlapping wording. Please merge them into a single paragraph to avoid redundancy.
  3. [Section 3.3] The term 'DAC discriminator' is used without prior introduction; please define it and cite the appropriate part of [7] (DAC / Improved RVQGAN).
  4. [Table 1 and Section 4.2.3] The ablation label 'w/o KL' in the text is inconsistent with the 'w/o KL-A' label in Table 1; please unify the naming.
  5. [Section 4.1] The subjective evaluation section would benefit from reporting the number of test samples per condition and whether the 9 listeners were experts or naive, as these details affect the interpretation of the MOS values.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the model is validated by held-out subjective and objective metrics plus ablations, and the co-authored citations are not load-bearing.

full rationale

The paper's derivation chain is not circular. The central claim is an empirical comparison: a trained CVAE model is evaluated on held-out subjective MOS (quality/naturalness/similarity) and objective metrics (PCC-E, RMSE-E, CER/WER), with all numbers reported in Table 1 rather than derived from fitted constants. The reported MOS values are listener ratings of generated samples, not quantities obtained by re-inserting optimized parameters, so there is no fitted-input-called-prediction pattern. The architecture borrows components from prior work (CVAE/VITS [5,11], XLS-R [18], FDRL [7]); some of those citations are co-authored by the present authors, but the paper does not invoke them as proof of its own success, and the ablations (w/o PP, w/ SEED, w/o KL-A) supply independent empirical grounding for the preprocessing, style-placement, and KL-annealing choices. The claimed linguistic-content preservation is supported by externally measured CER/WER against Whisper on the lingual subset, not by the definition of the input features. The internal contradiction between the abstract's 'outperformed baselines in ... naturalness' and Table 1's MOS-N 3.22 for Free-VC versus 3.16 for the proposed model is a correctness/reporting problem, not a circularity; likewise, the 9-participant panel without significance tests and the reader's concern that XLS-R features may not preserve non-linguistic expressive content are generalization and statistical-support concerns, not self-referential reductions. The paper itself flags f0 estimation limitations in Section 3.1, but that is a scope limitation, not a circular step. No equation or fitted parameter is renamed as a prediction, no uniqueness claim is imported from the authors' prior work, and no ansatz is smuggled in through self-citation. Therefore the appropriate circularity score is 0.

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

The paper introduces no new physical or conceptual entities; it relies on standard VC constructs (style vector, latent variables) and makes domain assumptions about the adequacy of speech-pretrained features and global style averaging for non-human sounds. The free parameters are training and preprocessing choices that shape the method but are not fitted to a closed-form result.

free parameters (4)
  • Information perturbation ranges = formant shift up to 1.8, pitch shift up to 3, pitch range 2
    Chosen by hand to generate timbre-perturbed waveforms that preserve content while accommodating wide frequency variations in non-human sounds (Section 3.1).
  • STFT hop length = 5 ms
    Design choice to capture fine temporal details in non-human sounds; not fitted to data but load-bearing for the preprocessing claim.
  • KL annealing step = 50,000
    Hyperparameter for cosine KL annealing schedule (Eq. 1); chosen to avoid KL-vanishing.
  • Loss weights = lambda_rec=45, lambda_fm=2, lambda_adv=1
    Training loss weights used in the generator loss; standard practice, but not justified by a principled derivation.
assumptions (3)
  • domain assumption XLS-R self-supervised speech features capture linguistic content independent of timbre for non-human vocalizations
    The linguistic encoder relies on the 12th-layer XLS-R representation from speech-trained Wav2Vec2.0 (Section 3.1); if these features do not encode content in screams, growls, or heavily perturbed audio, content preservation fails.
  • domain assumption An utterance-level averaged mel style vector sufficiently represents non-human timbres, including transient-rich growls
    The reference encoder averages a mel spectrogram over time to produce a global style vector (Section 3.1); transient and rapidly varying timbres may not be fully captured by a single average.
  • domain assumption The CVAE latent space trained on reconstruction disentangles style and content well enough for zero-shot conversion to extreme non-human timbres
    Inference combines source linguistic content with reference style from different utterances; this transfer is not guaranteed by the reconstruction training objective.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When Humans Growl and Birds Speak: High-Fidelity Voice Conversion from Human to Animal and Designed Sounds." pith.science (2026). https://pith.science/paper/YUBTNXOF

@misc{pith2026250524336,
  author       = {Pith},
  title        = {Pith review of: When Humans Growl and Birds Speak: High-Fidelity Voice Conversion from Human to Animal and Designed Sounds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YUBTNXOF}},
  note         = {Machine review of arXiv:2505.24336}
}
read the original abstract

Human to non-human voice conversion (H2NH-VC) transforms human speech into animal or designed vocalizations. Unlike prior studies focused on dog-sounds and 16 or 22.05kHz audio transformation, this work addresses a broader range of non-speech sounds, including natural sounds (lion-roars, birdsongs) and designed voice (synthetic growls). To accomodate generation of diverse non-speech sounds and 44.1kHz high-quality audio transformation, we introduce a preprocessing pipeline and an improved CVAE-based H2NH-VC model, both optimized for human and non-human voices. Experimental results showed that the proposed method outperformed baselines in quality, naturalness, and similarity MOS, achieving effective voice conversion across diverse non-human timbres. Demo samples are available at https://nc-ai.github.io/speech/publications/nonhuman-vc/

Figures

Figures reproduced from arXiv: 2505.24336 by the authors.

Figure 1
Figure 1. mel spectrogram visualizations including examples of (a) human speech, (b) human vocalization: yelling, (c) animal sound: birdsong, (d) artificailly designed sound. identifying individual birds or their emotional variations re￾mains difficult. Recent human VC studies [1, 2, 3, 4, 5] have optimized in￾put features for human speech, assuming stationarity by 40-80 frames and 10-20ms hop lengths while primarily modeling… view at source ↗
Figure 2
Figure 2. Proposed model architecture Therefore, our model achieved 3.16, 3.16, and 3.78 in quality, naturalness, and similarity MOS, respectively, and reached bet￾ter in most qualitative metrics, surpassing previous methods. The main contributions of this study are as follows: • Unlike prior work [6] limited to dog-sound conversion, our approach enables broader non-human voice conversion, in￾cluding natural and sound-designe… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 35 canonical work pages

  1. [1]

    When Humans Growl and Birds Speak: High-Fidelity Voice Conversion from Human to Animal and Designed Sounds

    Introduction Deep-learning based voice conversion (VC) has attained high- fidelity speech transformation [1, 2, 3, 4]. However, con- ventional VC methods cannot reliably synthesize non-speech vocalizations demanded by games, sci-fi, and interactive me- dia. Consequently, production pipelines still rely on hand- crafted, complex, time-intensive processes t...

    work page Pith review arXiv 2025

  2. [2]

    designed voices

    Related Works 2.1. V oice to Sound Effect Synthesis Several studies [8, 9] have explored voice to sound effect (SFX) generation. However, these approaches were evaluated solely on an SFX database with seven categories, and primarily fo- cused on generating SFX rather than preserving linguistic con- tent. Hence, they are not directly applicable to voice co...

    work page

  3. [3]

    Figure 2 presents an overview of the workflow of the system

    Proposed Method This section outlines the preprocessing and feature extrac- tion techniques used to accommodate human and non-human voices and improvements in the model architecture and training methodology. Figure 2 presents an overview of the workflow of the system. 3.1. Input Feature Set and Preprocessing Method Waveform, STFT and Mel: A sampling rate ...

    work page

  4. [4]

    w/o PP” indicates the proposed model using conventional (speech-focused) preprocessing method instead of our proposed pipeline, “w/ SEED

    Experiments 4.1. Experimental Settings We used internally collected dataset consisted of 82,008 au- dio samples, including the CORE 6 library from Pro Sound Effects [28]. The dataset was categorized as follows: (1) Ex- clamations and expressive utterances (37,332 samples), encom- passing vocal shouts, laughter, and similar expressions. (2) Sound-designed ...

    work page

  5. [5]

    It accu- rately captured transient-rich signals and wide-frequency–range of non-human sounds

    Conclusion The proposed H2NH-VC model effectively converted human speech into diverse non-human timbres, achieving significant MOS (Q/N/S) improvements over prior methods. It accu- rately captured transient-rich signals and wide-frequency–range of non-human sounds. The proposed method outcomes hint at its strong potential for use cases demanding extreme t...

    work page

  6. [6]

    Acknowledgements This research was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2024 (Project Name: Development of Co-Pilot technol- ogy for automatic completion of generative AI-based 3D Webtoon, Project Number: RS-2024-00400004, Contribution ...

    work page 2024

  7. [7]

    Neural anal- ysis and synthesis: Reconstructing speech from self-supervised representations,

    H. Choi, J. Lee, W. Kim, J. Lee, H. Heo, and K. Lee, “Neural anal- ysis and synthesis: Reconstructing speech from self-supervised representations,” in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 16 251–16 265

    work page 2021

  8. [8]

    Dddm-vc: Decoupled denoising dif- fusion models with disentangled representation and prior mixup for verified robust voice conversion,

    H. Choi, S. Lee, and S. Lee, “Dddm-vc: Decoupled denoising dif- fusion models with disentangled representation and prior mixup for verified robust voice conversion,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 16, 2024, pp. 17 862–17 870

    work page 2024

Show all 36 references
  1. [9]

    Diff-hiervc: Diffusion-based hierar- chical voice conversion with robust pitch generation and masked prior for zero-shot speaker adaptation,

    H.Choi, S.Lee, and S.Lee, “Diff-hiervc: Diffusion-based hierar- chical voice conversion with robust pitch generation and masked prior for zero-shot speaker adaptation,” in Proc. Interspeech , 2023, pp. 2283–2287

    2023

  2. [10]

    Hiervst: Hierarchical adap- tive zero-shot voice style transfer,

    S. Lee, H. Choi, H. Oh, and S. Lee, “Hiervst: Hierarchical adap- tive zero-shot voice style transfer,” inProc. Interspeech, 2023

    2023

  3. [11]

    FreeVC: Towards high-quality text- free one-shot voice conversion,

    J. Li, W. Tu, and L. Xiao, “FreeVC: Towards high-quality text- free one-shot voice conversion,” in2023 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP) , 2023, pp. 1–5

    2023

  4. [12]

    Speak like a dog: Human to non-human creature voice conversion,

    K. Suzuki, S. Sakamoto, T. Taniguchi, and H. Kameoka, “Speak like a dog: Human to non-human creature voice conversion,” in Asia-Pacific Signal and Information Processing Association An- nual Summit and Conference (APSIPA ASC) , 2022, pp. 1388– 1393

    2022

  5. [13]

    High-fidelity audio compression with improved rvqgan,

    R. Kumar, P. Seetharaman, A. Luebs, I. Kumar, and K. Kumar, “High-fidelity audio compression with improved rvqgan,” in Ad- vances in Neural Information Processing Systems, vol. 36, 2023, pp. 27 980–27 993

    2023

  6. [14]

    DDSP-SFX: Acoustically-guided sound ef- fects generation with differentiable digital signal processing,

    S. Lee and J. Smith, “DDSP-SFX: Acoustically-guided sound ef- fects generation with differentiable digital signal processing,” in Digital Audio Effects Conference (DAFx), 2024, pp. 216–221

    2024

  7. [15]

    T-FOLEY: A controllable waveform-domain diffusion model for temporal-event-guided fo- ley sound synthesis,

    Y . Chung, J. Lee, and J. Nam, “T-FOLEY: A controllable waveform-domain diffusion model for temporal-event-guided fo- ley sound synthesis,” in 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024

    2024

  8. [16]

    Dehumaniser2 - creature & monster sound design

    “Dehumaniser2 - creature & monster sound design.” [Online]. Available: https://www.krotosaudio.com/dehumaniser2/

  9. [17]

    Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,

    J. Kim, J. Kong, and J. Son, “Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,” in Pro- ceedings of the 38th International Conference on Machine Learn- ing, ser. Proceedings of Machine Learning Research, vol. 139. PMLR, 18–24 Jul 202...

    2021

  10. [18]

    Wave-tacotron: Spectrogram-free end-to-end text- to-speech synthesis,

    R. J. Weiss, R. Skerry-Ryan, E. Battenberg, S. Mariooryad, and D. P. Kingma, “Wave-tacotron: Spectrogram-free end-to-end text- to-speech synthesis,” in ICASSP 2021 - 2021 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 5679–5683

    2021

  11. [19]

    A systematic explo- ration of joint-training for singing voice synthesis,

    Y . Wu, Y . Yu, J. Shi, T. Qian, and Q. Jin, “A systematic explo- ration of joint-training for singing voice synthesis,” in 2024 IEEE 14th International Symposium on Chinese Spoken Language Pro- cessing (ISCSLP), 2024, pp. 289–293

    2024

  12. [20]

    Audioldm: Text-to-audio generation with latent diffusion models,

    H. Liu, Z. Chen, Y . Yuan, X. Mei, X. Liu, D. Mandic, W. Wang, and M. D. Plumbley, “Audioldm: Text-to-audio generation with latent diffusion models,” in Proceedings of the 40th International Conference on Machine Learning , vol. 202, 2023, pp. 21 450– 21 474

    2023

  13. [21]

    V AE with a vampprior,

    J. Tomczak and M. Welling, “V AE with a vampprior,” inProceed- ings of the Twenty-First International Conference on Artificial In- telligence and Statistics , ser. Proceedings of Machine Learning Research, A. Storkey and F. Perez-Cruz, Eds., vol. 84. PMLR, 09–11 Apr 2018, pp. ...

    2018

  14. [22]

    Flowtron: an autoregressive flow-based generative network for text-to- speech synthesis,

    R. Valle, K. J. Shih, R. Prenger, and B. Catanzaro, “Flowtron: an autoregressive flow-based generative network for text-to- speech synthesis,” in International Conference on Learning Representations, 2021. [Online]. Available: https://openreview. net/forum?id=Ig53hpHxS4

    2021

  15. [23]

    Meta-stylespeech: Multi-speaker adaptive text-to-speech generation,

    D. Min, D. Lee, E. Yang, and S. Hwang, “Meta-stylespeech: Multi-speaker adaptive text-to-speech generation,” inProceedings of the 38th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, vol. 139. PMLR, 18–24 Jul 2021, pp. 7748–7759

    2021

  16. [24]

    XLS-R: Self-supervised cross-lingual speech representation learning at scale,

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. von Platen, Y . Saraf, J. Pino, A. Baevski, A. Conneau, and M. Auli, “XLS-R: Self-supervised cross-lingual speech representation learning at scale,” in Proc. Interspeech , 2022, pp. 2278–2282. [Online]. A...

    2022

  17. [25]

    wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,” in Advances in Neural Information Processing Systems , vol. 33, 2020, pp. 12 449–12 460

    2020

  18. [26]

    Praat: doing phonetics by computer (Computer program),

    P. Boersma and D. Weenink, “Praat: doing phonetics by computer (Computer program),” Version 6.1.38, retrieved 2 January 2021 http://www.praat.org/, 2021

    2021

  19. [27]

    Yet another algorithm for pitch tracking,

    K. Kasi and S. A. Zahorian, “Yet another algorithm for pitch tracking,” in 2002 IEEE International Conference on Acoustics, Speech, and Signal Processing, vol. 1, 2002, pp. I–361–I–364

    2002

  20. [28]

    Crepe: A convolu- tional representation for pitch estimation,

    J. Kim, J. Salamon, P. Li, and J. P. Bello, “Crepe: A convolu- tional representation for pitch estimation,” in 2018 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, pp. 161–165

    2018

  21. [29]

    SPICE: Self-supervised pitch estimation,

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

    2020

  22. [30]

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

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

    2023

  23. [31]

    Autoencoding beyond pixels using a learned similarity metric,

    A. B. L. Larsen, S. K. Sønderby, H. Larochelle, and O. Winther, “Autoencoding beyond pixels using a learned similarity metric,” in Proceedings of the 33rd International Conference on Ma- chine Learning, ser. Proceedings of Machine Learning Research, vol. 48, 2016, pp. 1558–1566

    2016

  24. [32]

    Least squares generative adversarial networks,

    X. Mao, Q. Li, H. Xie, R. Y . K. Lau, Z. Wang, and S. P. Smolley, “Least squares generative adversarial networks,” in 2017 IEEE International Conference on Computer Vision (ICCV) , 2017, pp. 2813–2821

    2017

  25. [33]

    Learning discourse-level di- versity for neural dialog models using conditional variational au- toencoders,

    T. Zhao, R. Zhao, and M. Eskenazi, “Learning discourse-level di- versity for neural dialog models using conditional variational au- toencoders,” in Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, 2017, pp. 654–664

    2017

  26. [34]

    Pro sound effects - core 6 library

    “Pro sound effects - core 6 library.” [Online]. Available: https://www.prosoundeffects.com/core-6/pro

  27. [35]

    Robust speech recognition via large-scale weak su- pervision,

    A. Radford, J. Kim, T. Xu, G. Brockman, C. Mcleavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,” in Proceedings of the 40th International Conference on Machine Learning, ser. Proceedings of Machine Learning Re- search, vol. 202. PMLR, 23–2...

    2023

  28. [36]

    HiFi-GAN: Generative adversarial networks for efficient and high fidelity speech synthesis,

    J. Kong, J. Kim, and J. Bae, “HiFi-GAN: Generative adversarial networks for efficient and high fidelity speech synthesis,” in Ad- vances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, Eds., vol. 33. Curran Associates,...

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.