Pith. sign in

REVIEW 3 major objections 4 minor 37 references

First Steps Towards Voice Anonymization for Code-Switching Speech

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A multilingual ASR-to-speech anonymizer is the first voice-privacy method shown to work on code-switched Mandarin-English and Spanish-English speech.

desk verdict First code-switching voice anonymization benchmark with a credible privacy result, but the utility claim is undercut by a same-ASR circularity and the paper's own analysis shows substantial code-switching loss. read the letter →

arxiv 2507.01765 v1 pith:C3ARHHCR submitted 2025-07-02 eess.AS

Sarina Meyer , Ekaterina Kolos , Ngoc Thang Vu This is my paper
classification eess.AS
keywords voiceanonymizationcode-switchingspeakerprivacymultilingualTTSASR-to-TTScascademixederrorrateequalcorpora
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

Voice anonymization usually hides a speaker's identity only in clean, read, monolingual English audio. This paper asks whether anonymizers still work when a speaker switches between Mandarin and English or Spanish and English within a conversation, and it prepares new evaluation splits from two code-switching corpora to find out. It compares a signal-processing method, a self-supervised neural method designed to be language-independent, and a multilingual ASR-to-TTS cascade adapted by the authors to switch languages word by word. The central result is that only the multilingual cascade both protects the speaker, with speaker-verification equal error rates near 50 percent, and preserves content, with only 1-6 percent absolute mixed-error-rate degradation; the two language-independent methods leave speech identifiable almost as well as the original audio. A secondary finding is that the multilingual ASR used for evaluation itself reduces code-switching during transcription, so utility numbers on this data mix recognition errors with anonymization errors.

What carries the argument

The carrying mechanism is the adapted GANmulti cascade: Whisper converts the input audio to text and identifies its language; a word-level language detector labels each word; and a multilingual FastSpeech2 text-to-speech system synthesizes the anonymized audio with the language embedding chosen per phone instead of once per utterance, so a switch inside an utterance can be reproduced. Speaker identity is hidden because an ECAPA-TDNN-like speaker representation is sampled from a GAN rather than taken from the true speaker. The two baselines fail, in this account, because they treat speech as a single language-agnostic signal: spectral shifting changes timbre without hiding identity, and the self-supervised content-unit system transfers the content but keeps speaker information recognizable.

What would settle it

Run the anonymized and original code-switching utterances through a code-switching-specific ASR (for example, Whisper adapted on SEAME) and recompute MER and code-switching-point counts; if the gap between original and anonymized MER shrinks or the reported code-switching reduction disappears, then the paper's utility and code-switching-preservation findings are artifacts of the evaluator rather than properties of the anonymizer.

Watch

Extended reading notes

Core claim

The paper's claim is that code-switching speech is anonymizable, but only if the anonymizer is explicitly multilingual. On SEAME (Mandarin-English) and MIAMI (Spanish-English), the adapted GANmulti model raises the equal error rate of a speaker verification attacker trained on anonymized data to 48-51 percent, close to the chance level of 50 percent, while increasing mixed error rate by only 1-6 percent absolute across language settings. The two language-independent baselines, one based on spectral shifting and one on self-supervised soft content units, reach equal error rates below 10 percent on almost all settings, comparable to no anonymization, and degrade utility by 13-42 percent MER. The authors also find that speaker verification is slightly easier on code-switched utterances than monolingual ones on original audio, and that the Whisper model used for utility evaluation transcribes code-switched utterances with around 39 percent MER, translating or dropping parts of one language and thereby reducing measured code-switching points after anonymization.

Load-bearing premise

The chain depends on Whisper-large-v3 being a fair transcriber of code-switched speech: the same model feeds the anonymizer its text and then scores the output, so if Whisper systematically translates, drops, or mislabels one of the languages, both the measured utility and the measured loss of code-switching points inherit that bias.

Editorial extensions

If this is right

  • If GANmulti's results hold, the practical route to anonymizing code-switched audio is an ASR-to-TTS cascade that re-synthesizes according to detected languages, not signal modification or language-agnostic representations.
  • Privacy protection at roughly 50 percent EER generalizes from English read speech to spontaneous conversational speech in two language pairs, so multilingual cascades are not limited to clean single-language data.
  • Utility evaluation on code-switched data needs its own yardstick: MER computed by a single-language model overstates anonymization harm because the recognizer itself translates or omits languages.
  • Code-switching behavior is only partially preserved: on SEAME 30 percent of utterances keep the same number of code-switching points after anonymization, and on MIAMI 26 percent; much of the measured MER increase tracks the loss of code-switching points.
  • The datasets and code released with the paper let other groups run the same privacy and utility comparison on additional code-switching language pairs.

Reading between the lines

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

  • Beyond the paper: because the GANmulti pipeline's input transcript and its utility metric both come from Whisper-large-v3, the favorable utility numbers could partly be an artifact of evaluating with the same recognizer that supplies the content; a code-switching-aware evaluator or human transcription judgment would separate recognition bias from anonymization degradation.
  • Beyond the paper: the observed 20-30 percent MER on original monolingual utterances suggests the corpora themselves, not just the methods, challenge current ASR; improving recognition of spontaneous, accented, code-switched speech would likely improve measured anonymization utility even without changing the anonymizer.
  • A testable extension would be to anonymize the same code-switched audio with GANmulti while prompting Whisper with each language explicitly, or with a code-switching-adapted recognizer, and then compare both code-switching-point retention and MER; this would quantify how much of the reported 1-6 percent degradation is due to the recognizer's language bias.
  • If privacy and utility results transfer to other language pairs, code-switching corpora could become a standard stress test for voice anonymizers, since they combine spontaneous speech, accent variation, and in-utterance language shifts in one benchmark.
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

3 major / 4 minor

Summary. The paper presents the first investigation of voice anonymization for code-switching speech. The authors prepare evaluation data from SEAME (Mandarin-English) and MIAMI (Spanish-English), evaluate three anonymization systems (the McAdams-coefficient baseline B2, the SSL5 self-supervised method, and a multilingual ASR-TTS cascade GANmulti which they adapt for code-switching), and report privacy (EER) and utility (MER/PER) scores. They find that only GANmulti achieves strong privacy (EER close to 50%) and low utility degradation (1-6% absolute MER increase), while the language-independent baselines fail. They also analyze Whisper's recognition of code-switched speech, showing high baseline MER and significant loss of code-switching points during anonymization.

Significance. If the results are confirmed, this is a valuable first step: it establishes a benchmark for code-switching voice anonymization and provides evidence that multilingual ASR-TTS cascades are currently the only viable approach, while signal-processing and self-supervised methods that ignore language structure fail. The paper is commendable for releasing code and data, for evaluating a semi-informed ASV attacker including finetuning on anonymized SEAME data (footnote 10), and for honestly documenting Whisper's limitations on code-switched spontaneous speech. However, the utility claim is confounded by the use of Whisper both inside GANmulti and as the evaluation metric, and the paper's own analysis shows substantial loss of code-switching after anonymization, so the headline conclusion is not yet fully established.

major comments (3)
  1. [Section 2.4 and Section 3.2] The utility evaluation is circular: Whisper-large-v3 generates the transcript that GANmulti synthesizes (Section 2.2) and also computes the MER (Section 2.3). The reported 1-6% absolute MER increase therefore largely measures how well GANmulti preserves the content of Whisper's own transcription, not how well it preserves the original spoken content. The paper's own results in Table 4 show that 65-71% of code-switching utterances have fewer code-switching points after anonymization, and 30-64% lose code-switching entirely; Section 4.2 documents translations, omissions, and near-homophone substitutions. To support the conclusion that GANmulti preserves utility for code-switching speech, the authors should re-score the anonymized audio using an independent ASR or human transcription based on gold transcripts, and report content preservation metrics (e.g., code-switching point retention, semantic similarity) rather than only aggregate MER.
  2. [Section 1 and Table 4] The paper states as a goal that the code-switching behavior of the original speaker should be preserved (Section 1). Table 4 directly contradicts this for GANmulti: in MIAMI, 64% of code-switching utterances lose code-switching entirely after anonymization, and in SEAME 30% do; even when some points remain, the average CSP drops from 3.47 to 1.75 (SEAME) and 1.47 to 0.52 (MIAMI). The MER numbers in Table 3 therefore understate the content degradation, and the claim that GANmulti preserves utility for code-switching speech is not established without a more fine-grained analysis of what is preserved.
  3. [Section 2.3] The phone error rate (PER) is computed using the IMS Toucan phonemizer (Section 2.3), which is also part of the GANmulti synthesis front-end. The PER is therefore not an independent confirmation of utility, contrary to the implication in Section 3.2 that it 'confirms these trends'. An independent phonemizer or a human-perceptual evaluation would be needed to break this dependence.
minor comments (4)
  1. [Section 2.4] The language-prompt settings for Whisper are tuned on the same evaluation data (SEAME and MIAMI), which may optimistically bias the reported MER numbers. Please clarify whether the prompt selection was done on a held-out development set (a dev split exists for SEAME) and, if not, report the sensitivity of the results to this choice.
  2. [Table 3] The column label 'Orig↓' is confusing because 'Orig' is not a direction; consider using 'Original' as a separate column header and keeping the arrows only for the metric columns.
  3. [Section 2.2] The text alternates between 'GANmulti' and 'GAN multi'; please use a single consistent spelling throughout.
  4. [Section 4.3] The subset-removal analysis in Section 4.3 would benefit from a clearer description of how overlapping subsets are handled and whether the reported MER changes are statistically tested; the current description ('we can estimate the effect') is vague.

Circularity Check

1 steps flagged · score 4.0 of 10

GANmulti's utility claim rests on a Whisper evaluation loop: the ASR that writes the transcript for synthesis is also the ASR that scores MER, so code-switching content loss can be invisible.

  1. other [Sections 2.2, 2.3, 2.4 and 3.2 (Table 3); supported by Sections 4.1 and 4.2]
    "Following [12], we use Whisper-large-v3 for this purpose, similar to its use in the GANmulti model. [...] Whisper is used for ASR in the anonymization of GANmulti as well as the utility evaluation."

    GANmulti synthesizes the anonymized audio from Whisper's transcript of the original speech, and the same Whisper-large-v3 model then computes the MER used to claim low utility degradation. Any systematic code-switching error made by Whisper (translations, omissions, near-homophones, as documented in Section 4.2) is baked into the resynthesized audio and re-recognized with the same bias, so that error cancels in the before/after MER comparison. The reported 1-6% absolute MER increase measures how consistently Whisper recognizes its own transcript after TTS resynthesis, not how much original spoken content is preserved.

full rationale

The paper contains no derivation-style circularity: it is an empirical evaluation of existing anonymization systems, and the central privacy result is measured with an independent ECAPA-TDNN attacker trained on LibriSpeech. However, the utility claim for GANmulti is partially circular in an evaluation sense. Whisper-large-v3 is both the front-end that generates the transcript GANmulti speaks and the evaluator that scores the anonymized audio's MER, and the paper explicitly states that Whisper is used for ASR in the anonymization as well as the utility evaluation. Since GANmulti's content is defined by Whisper's own transcription of the original, the before/after MER comparison can hide content loss that Whisper reproduces consistently. The authors are transparent about this limitation and even call for more diverse evaluation frameworks; the self-disclosure lowers the score, but the headline conclusion that only GANmulti achieves low utility degradation is not fully independent of the evaluator. No other circular steps, uniqueness-importation, or ansatz-smuggling via citation were found.

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

The central claims rest on evaluation choices and domain assumptions rather than on fitted model constants. The Whisper prompt settings, the special-character suppression, the duration threshold and the speaker filter are all hand-chosen and affect the reported numbers. The paper also assumes that preserving code-switching behavior is desirable, that an English-trained ASV measures privacy in code-switching data, and that Whisper can serve simultaneously as anonymization front-end and utility evaluator. No new physical or theoretical entities are introduced.

free parameters (4)
  • Whisper language prompt setting per dataset = SEAME: 'Mandarin'; MIAMI: auto-detect with Spanish re-run
    Section 2.4: the prompt was chosen experimentally to give best ASR results on each dataset. This choice affects both the transcript used for GANmulti synthesis and the MER utility metric, and it is a free parameter tuned to the data.
  • Special-character probability suppression threshold = Not specified ('set the output probabilities ... to a low value')
    Section 2.4: a hand-chosen threshold that changes how Whisper transcribes digits, currency and punctuation. The exact value is not reported, so the behavior is not fully reproducible.
  • Short-utterance concatenation duration threshold = 2 seconds
    Section 2.1: utterances shorter than 2 seconds are concatenated with other short utterances from the same speaker. This changes the unit of evaluation and the statistics in Tables 1 and 3.
  • Minimum utterance count for MIAMI speaker inclusion = 20 utterances per language setting
    Section 2.1: speakers with fewer than 20 utterances in each language setting are excluded, reducing the MIAMI test set to 25 speakers and potentially biasing the evaluated population.
assumptions (4)
  • domain assumption The code-switching behavior of the original speaker should be preserved during anonymization.
    Section 1: 'we assume that the code-switching behavior of the original speaker should be preserved.' This is a design choice, not an empirical result, and it drives the utility definition.
  • domain assumption A semi-informed ASV attacker trained on English LibriSpeech transfers to Mandarin-English and Spanish-English code-switching speech.
    Section 2.3: the ASV model is trained on LibriSpeech train-clean-360, with a finetuning experiment on SEAME train data that showed no improvement. The attack on MIAMI is not finetuned at all, so cross-domain transfer is assumed.
  • domain assumption Whisper-large-v3 is a valid ASR for both anonymization and utility measurement of code-switching speech.
    Section 2.4: Whisper is used inside GANmulti and as the MER/CER evaluator. The paper acknowledges Whisper expects single-language audio and does not have full code-switching support, so this assumption is load-bearing for the utility analysis.
  • domain assumption Only intra-sentential code-switching counts as code-switching; inter-sentential switches are treated as monolingual.
    Section 2.1: the authors split utterances into EN, ZH/ES and CS, labeling only intra-sentential switches as CS. This changes the composition of the CS subset and thus all reported CS results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of First Steps Towards Voice Anonymization for Code-Switching Speech." pith.science (2026). https://pith.science/paper/C3ARHHCR

@misc{pith2026250701765,
  author       = {Pith},
  title        = {Pith review of: First Steps Towards Voice Anonymization for Code-Switching Speech},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C3ARHHCR}},
  note         = {Machine review of arXiv:2507.01765}
}
read the original abstract

The goal of voice anonymization is to modify an audio such that the true identity of its speaker is hidden. Research on this task is typically limited to the same English read speech datasets, thus the efficacy of current methods for other types of speech data remains unknown. In this paper, we present the first investigation of voice anonymization for the multilingual phenomenon of code-switching speech. We prepare two corpora for this task and propose adaptations to a multilingual anonymization model to make it applicable for code-switching speech. By testing the anonymization performance of this and two language-independent methods on the datasets, we find that only the multilingual system performs well in terms of privacy and utility preservation. Furthermore, we observe challenges in performing utility evaluations on this data because of its spontaneous character and the limited code-switching support by the multilingual speech recognition model.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 27 canonical work pages

  1. [1]

    Common approaches modify the audio using signal processing [1], voice conversion [2, 3, 4], or cascade systems of speech recognition (ASR) and text-to- speech (TTS) [5]

    Introduction V oice anonymization (V A) refers to the task of hiding the iden- tity of a speaker in an audio. Common approaches modify the audio using signal processing [1], voice conversion [2, 3, 4], or cascade systems of speech recognition (ASR) and text-to- speech (TTS) [5]. The main goal, as defined by the V oice Pri- vacy Challenges (VPC) [6, 7, 8],...

    work page 2024

  2. [2]

    First Steps Towards Voice Anonymization for Code-Switching Speech

    Methods 2.1. Data We perform all experiments on two datasets, S EAME (Mandarin-English) and M IAMI (Spanish-English). They con- tain two types of code-switching: Inter-sentential, in which the language switch happens between utterances, and intra- sentential with language switches within an utterance. We split the data into single utterances and denote on...

    work page Pith review arXiv 2025

  3. [3]

    Anonymization of VPC datasets In order to understand the general anonymization abilities of each V A model, we first test them on the English datasets provided by the VPC [7]

    Experiments 3.1. Anonymization of VPC datasets In order to understand the general anonymization abilities of each V A model, we first test them on the English datasets provided by the VPC [7]. The results are shown in Table 2. All methods increase privacy protection except SSL on Lib- riSpeech. Generally, the anonymization of B2 and SSL re- sult in only w...

    work page

  4. [4]

    jobs” and “jumps

    Analysis Among the V A systems we tested, only GANmulti could achieve a high level of privacy and keep the utility degradation compa- rably low. However, the utility scores on the original data are substantially higher than on the standard VPC data. Thus, we examine the outcome of the ASR before and after anonymiza- tion with GAN multi more closely to und...

    work page

  5. [5]

    We prepare two Mandarin-English and Spanish-English datasets for the evalua- tion and experiment with three anonymization methods

    Conclusion In this paper, we present the first work of applying voice anonymization to code-switching speech. We prepare two Mandarin-English and Spanish-English datasets for the evalua- tion and experiment with three anonymization methods. Two of them are designed as language-independent but fail to achieve sufficient privacy preservation even on monolin...

    work page

  6. [6]

    Acknowledgements This work is funded by the Deutsche Forschungsgemein- schaft (DFG, German Research Foundation) – Project: Mul- tilingual Controllable V oice Privacy (V oiPy) - Project number 533241795

  7. [7]

    Speaker anonymisation using the mcadams coefficient,

    J. Patino, N. Tomashenko, M. Todisco, A. Nautsch, and N. Evans, “Speaker anonymisation using the mcadams coefficient,” inInter- speech 2021, 2021, pp. 1099–1103

    work page 2021

  8. [8]

    Language-Independent Speaker Anonymization Approach Us- ing Self-Supervised Pre-Trained Models,

    X. Miao, X. Wang, E. Cooper, J. Yamagishi, and N. Tomashenko, “Language-Independent Speaker Anonymization Approach Us- ing Self-Supervised Pre-Trained Models,” in Proc. The Speaker and Language Recognition Workshop (Odyssey 2022) , 2022, pp. 279–286

    work page 2022

Show all 37 references
  1. [9]

    Speaker Anonymization Using Neural Audio Codec Language Models,

    M. Panariello, F. Nespoli, M. Todisco, and N. Evans, “Speaker Anonymization Using Neural Audio Codec Language Models,” in ICASSP 2024 - 2024 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP) , 2024, pp. 4725– 4729

    2024

  2. [10]

    Differ- entially private speaker anonymization,

    A. S. Shamsabadi, B. M. L. Srivastava, A. Bellet, N. Vauquier, E. Vincent, M. Maouche, M. Tommasi, and N. Papernot, “Differ- entially private speaker anonymization,”Proc. Privacy Enhancing Technologies, vol. 2023, no. 1, pp. 98–114, Jan. 2023

    2023

  3. [11]

    Prosody Is Not Identity: A Speaker Anonymization Approach Using Prosody Cloning,

    S. Meyer, F. Lux, J. Koch, P. Denisov, P. Tilli, and N. T. Vu, “Prosody Is Not Identity: A Speaker Anonymization Approach Using Prosody Cloning,” in ICASSP 2023 - 2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  4. [12]

    Introducing the V oicePrivacy Initiative,

    N. Tomashenko, B. M. L. Srivastava, X. Wang, E. Vincent, A. Nautsch, J. Yamagishi, N. Evans, J. Patino, J.-F. Bonastre, P.-G. No´e, and M. Todisco, “Introducing the V oicePrivacy Initiative,” in Interspeech 2020, 2020, pp. 1693–1697

    2020

  5. [13]

    The V oicePrivacy 2022 Challenge: Progress and Perspec- tives in V oice Anonymisation,

    M. Panariello, N. Tomashenko, X. Wang, X. Miao, P. Champion, H. Nourtel, M. Todisco, N. Evans, E. Vincent, and J. Yamag- ishi, “The V oicePrivacy 2022 Challenge: Progress and Perspec- tives in V oice Anonymisation,” IEEE/ACM Transactions on Au- dio, Speech, and Language Proces...

    2022

  6. [14]

    The V oicePrivacy 2024 Challenge Evaluation Plan,

    N. Tomashenko, X. Miao, P. Champion, S. Meyer, X. Wang, E. Vincent, M. Panariello, N. Evans, J. Yamagishi, and M. Todisco, “The V oicePrivacy 2024 Challenge Evaluation Plan,”

    2024

  7. [15]

    Code-switching in automatic speech recognition: The issues and future directions,

    M. B. Mustafa, M. A. Yusoof, H. K. Khalaf, A. A. Rahman Mah- moud Abushariah, M. L. M. Kiah, H. N. Ting, and S. Muthaiyah, “Code-switching in automatic speech recognition: The issues and future directions,” Applied Sciences , vol. 12, no. 19, p. 9541, 2022

    2022

  8. [16]

    Lib- rispeech: An ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: An ASR corpus based on public domain audio books,” in 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015, pp. 5206–5210

    2015

  9. [17]

    CSTR VCTK Cor- pus: English Multi-speaker Corpus for CSTR V oice Cloning Toolkit (version 0.92),

    J. Yamagishi, C. Veaux, and K. MacDonald, “CSTR VCTK Cor- pus: English Multi-speaker Corpus for CSTR V oice Cloning Toolkit (version 0.92),” 2019

    2019

  10. [18]

    IEMOCAP: interac- tive emotional dyadic motion capture database,

    C. Busso, A. Kazemzadeh, and C.-C. Lee, “IEMOCAP: interac- tive emotional dyadic motion capture database,” 2008

    2008

  11. [19]

    Probing the feasibility of mul- tilingual speaker anonymization,

    S. Meyer, F. Lux, and N. T. Vu, “Probing the feasibility of mul- tilingual speaker anonymization,” in Interspeech 2024, 2024, pp. 4448–4452

    2024

  12. [20]

    Musa: Multi-lingual speaker anonymization via serial disentan- glement,

    J. Yao, Q. Wang, P. Guo, Z. Ning, Y . Yang, Y . Pan, and L. Xie, “Musa: Multi-lingual speaker anonymization via serial disentan- glement,” IEEE Transactions on Audio, Speech and Language Processing, vol. 33, pp. 1664–1674, 2025

    2025

  13. [21]

    Analyzing Language-Independent Speaker Anonymization Framework under Unseen Conditions,

    X. Miao, X. Wang, E. Cooper, J. Yamagishi, and N. Tomashenko, “Analyzing Language-Independent Speaker Anonymization Framework under Unseen Conditions,” in Interspeech 2022 , 2022, pp. 4426–4430

    2022

  14. [22]

    HiFi-GAN: Generative Adversar- ial Networks for Efficient and High Fidelity Speech Synthesis,

    J. Kong, J. Kim, and J. Bae, “HiFi-GAN: Generative Adversar- ial Networks for Efficient and High Fidelity Speech Synthesis,” in Advances in Neural Information Processing Systems , vol. 33. Curran Associates, Inc., pp. 17 022–17 033

  15. [23]

    From codeswitching via language mixing to fused lects: Toward a dynamic typology of bilingual speech,

    P. Auer, “From codeswitching via language mixing to fused lects: Toward a dynamic typology of bilingual speech,” International Journal of Bilingualism, vol. 3, no. 4, pp. 309–332, 1999

    1999

  16. [24]

    Seame: a mandarin- english code-switching speech corpus in south-east asia,

    D.-C. Lyu, T.-P. Tan, E. S. Chng, and H. Li, “Seame: a mandarin- english code-switching speech corpus in south-east asia,” inInter- speech 2010, 2010, pp. 1986–1989

    2010

  17. [25]

    Spectral fusion, spectral parsing and the forma- tion of auditory images,

    S. McAdams, “Spectral fusion, spectral parsing and the forma- tion of auditory images,” Ph.D. dissertation, Stanford University, Stanford, California, 05/1984 1984

    1984

  18. [26]

    Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,”IEEE/ACM Trans. Audio, Speech and Lang. Proc., vol. 29, p. 3451–3460, Oct. 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]

    Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,” in Interspeech 2020, 2020, pp. 3830–3834

    2020

  21. [29]

    V oicePAT: An Efficient Open- Source Evaluation Toolkit for V oice Privacy Research,

    S. Meyer, X. Miao, and N. T. Vu, “V oicePAT: An Efficient Open- Source Evaluation Toolkit for V oice Privacy Research,” IEEE Open Journal of Signal Processing, vol. 5, pp. 257–265, 2024

    2024

  22. [30]

    Generative adver- sarial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde- Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adver- sarial nets,” in Advances in Neural Information Processing Sys- tems, Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K. Weinberger, Eds., vol....

    2014

  23. [31]

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

    A. Radford, J. W. 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. ICML’23. JMLR.org, 2023

    2023

  24. [32]

    Fastspeech 2: Fast and high-quality end-to-end text to speech,

    Y . Ren, C. Hu, X. Tan, T. Qin, S. Zhao, Z. Zhao, and T.-Y . Liu, “Fastspeech 2: Fast and high-quality end-to-end text to speech,” in International Conference on Learning Representations, 2021

    2021

  25. [33]

    The ims tou- can system for the blizzard challenge 2021,

    F. Lux, J. Koch, A. Schweitzer, and N. Thang Vu, “The ims tou- can system for the blizzard challenge 2021,” inThe Blizzard Chal- lenge 2021, 2021, pp. 14–19

    2021

  26. [34]

    Anonymizing speech with generative adversarial networks to preserve speaker privacy,

    S. Meyer, P. Tilli, P. Denisov, F. Lux, J. Koch, and N. T. Vu, “Anonymizing speech with generative adversarial networks to preserve speaker privacy,” in 2022 IEEE Spoken Language Tech- nology Workshop (SLT), 2023, pp. 912–919

    2022

  27. [35]

    Meta Learning Text-to-Speech Synthesis in over 7000 Languages,

    F. Lux, S. Meyer, L. Behringer, F. Zalkow, P. Do, M. Coler, E. A. P. Habets, and N. T. Vu, “Meta Learning Text-to-Speech Synthesis in over 7000 Languages,” inInterspeech 2024. ISCA, 2024

    2024

  28. [37]

    Adapt- ing openai’s whisper for speech recognition on code-switch mandarin-english seame and asru2019 datasets,

    Y . Yang, Y . Peng, H. Huang, E. S. Chng, and X. Zhong, “Adapt- ing openai’s whisper for speech recognition on code-switch mandarin-english seame and asru2019 datasets,” in2024 Asia Pa- cific Signal and Information Processing Association Annual Sum- mit and Conference (APSIPA ...

    2024

  29. [2024]

    Available: https://arxiv.org/abs/2404.02677

    [Online]. Available: https://arxiv.org/abs/2404.02677

    arXiv

Pith tools

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