Pith. sign in

REVIEW 4 major objections 7 minor 31 references

Developing multilingual speech synthesis system for Ojibwe, Mi'kmaq, and Maliseet

T0 review · 4 major / 7 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Training a single flow-matching TTS model on three genetically related Algonquian languages—Ojibwe, Mi'kmaq, and Maliseet—outperforms separate monolingual models, especially when data are scarce, and attention-free layers match…

desk verdict First TTS for Mi'kmaq and Maliseet with an honest efficiency comparison, but the multilingual improvement claim is confounded by data volume and oversampling. read the letter →

arxiv 2502.02703 v1 pith:6WK4SFSE submitted 2025-02-04 cs.CL cs.AIcs.LGcs.SDeess.AS

Shenran Wang , Changbing Yang , Mike Parkhill , Chad Quinn , Christopher Hammerly , Jian Zhu This is my paper · ORCID
classification cs.CLcs.AIcs.LGcs.SDeess.AS
keywords text-to-speechflowmatchingmultilingualTTSlow-resourcelanguagesIndigenouslanguagerevitalizationOjibweMi'kmaqMaliseet
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 develops text-to-speech (TTS) systems for three Indigenous Algonquian languages of Canada—Ojibwe, Mi'kmaq, and Maliseet—using a single flow-matching model trained on all three languages at once. The central claim is that training one multilingual model on related languages produces better synthetic voices than training a separate model on each language alone, and that the benefit is largest for the languages with the least data. The paper also shows that replacing self-attention with cheaper attention-free layers (Mamba2, Hydra, FNet) keeps quality nearly unchanged while roughly halving batched memory use. If these results hold, builders of speech tools for endangered languages can get usable voices from a few hours of home-recorded audio per language and can deploy them on ordinary computers.

What carries the argument

The load-bearing object is a multilingual Matcha-TTS model: a conditional flow-matching decoder that predicts mel spectrograms from text, augmented with learned speaker and language embeddings concatenated to the text-encoder output. Flow matching is a generative method that trains a network to follow a vector field transporting noise to data; here it generates the mel spectrogram in about 10 denoising steps, conditioned on the text and the chosen voice. The architecture comparison works by replacing the self-attention layers with three attention-free sequence mixers: Mamba2 selective state-space blocks, Hydra bidirectional quasiseparable mixers, and an FNet-style Fourier layer. Because these mixers have subquadratic or linear complexity, they reduce memory use at equal model scale, and the paper measures their effect on quality, throughput, and memory.

What would settle it

Train the same architecture twice with identical total data and speaker balance, once with the true language labels and once with the language embeddings randomly permuted across speakers. If the permuted model matches the true multilingual model's improvement over monolingual baselines, the claim that shared Algonquian structure drives the gain is refuted; a complementary check uses three typologically unrelated languages at matched data sizes.

Watch

Extended reading notes

Core claim

The paper's central discovery is practical: for four speakers of three genetically related Algonquian languages, one flow-matching TTS model trained on all available speech (about 23 hours, oversampled so each speaker contributes comparable duration) produces voices that generally score as well as or better than a self-attention model trained on each speaker's data alone. The clearest gains appear for the smallest corpora—Mi'kmaq at 2h23m and one Ojibwe speaker at 1h41m—where the multilingual models reduce log-amplitude RMSE and mel-cepstral distortion and improve several intelligibility metrics. The authors attribute this to the model learning shared phonological structure across related languages while conditioning on learned speaker and language embeddings, and they take it as evidence that diversity across related languages can partly compensate for lack of data in any one language. Swapping self-attention for attention-free sequence mixers (Mamba2, Hydra, FNet) produces close quality, with Hydra the nearest to self-attention, while cutting batched memory from 4.6GB to about 2.3–2.5GB. The paper also reports that standard MOS evaluation proved unreliable in this context, with one Ojibwe rater rating every voice, natural or synthetic, at the top of the scale, which the authors read as a cultural gap in the evaluation protocol.

Load-bearing premise

The comparison assumes that the multilingual model's advantage comes from the three languages being related, rather than simply from being trained on far more total audio (about 23 hours versus 1.7–11.8 hours per speaker); no control matches data volume or uses unrelated languages.

Editorial extensions

If this is right

  • Future data collection for endangered-language TTS can prioritize gathering smaller corpora from several related languages over one large corpus, since the multilingual model exploits shared structure to compensate for scarce data.
  • Attention-free architectures, particularly Hydra, are viable substitutes for self-attention in low-resource TTS, roughly halving batched memory with little quality loss, which matters for deployment on common devices.
  • Human evaluation of synthetic voices for Indigenous communities should be redesigned with community cultural norms in mind, because the standard MOS rating task produced unusable ratings from at least one participant.
  • A usable synthetic voice can be built from as little as roughly 1.7 hours of home-recorded speech for a single speaker when related-language data are available for joint training.

Reading between the lines

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

  • Our inference, not the paper's: the multilingual advantage is overdetermined, because the multilingual model also sees about 23 hours of audio while each monolingual model sees only 1.7–11.8 hours; a matched-volume control is needed before attributing the improvement to Algonquian typological similarity.
  • Our inference: the oversampling procedure equalizes each speaker's contribution, so the multilingual model may benefit mainly from more stable, more balanced training rather than from cross-lingual phonetic transfer.
  • Our inference: if the mechanism is genuine transfer among related languages, then adding a small corpus in a sister language should help a low-resource voice more than adding the same hours of an unrelated language; this is testable with the released code.
  • Our inference: the home-recording protocol trades studio cleanliness for speaker comfort and community control, and the paper's results suggest ambient noise is not a blocker; however, the paper does not quantify how much noise the model tolerates, so a systematic degradation study would sharpen deployment guidance.
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 / 7 minor

Summary. The paper develops lightweight flow-matching multilingual TTS systems for Ojibwe, Mi'kmaq, and Maliseet, using community-centered recordings and a character-based tokenizer. It extends Matcha-TTS with speaker and language embeddings and compares the standard self-attention architecture against three attention-free sequence mixers: FNet, Mamba2, and Hydra. The main empirical claims are that (i) multilingual training on typologically related Algonquian languages outperforms per-speaker monolingual training, especially for low-resource settings, and (ii) attention-free architectures are competitive with self-attention while being more memory-efficient. The paper also discusses the cultural limitations of standard MOS-style human evaluation for Indigenous communities.

Significance. If the multilingual-transfer claim were established, the paper would be a useful contribution to low-resource and Indigenous-language TTS, with practical guidance for data collection and deployment. The authors provide reproducible artifacts: code is publicly available, training details are tabulated, and the community-centered recording protocol is described carefully. The ethical discussion of evaluation protocols is a genuine strength and is relevant well beyond this specific system. However, the headline claim is currently not supported by the experimental design because the multilingual models receive much more data and more target-speaker exposure than the monolingual models, and the architecture comparison is also confounded by differing hyperparameters. The value of the paper is therefore conditional on additional controls or a substantially weakened claim.

major comments (4)
  1. [Section 2 / Table 2] The central comparison between monolingual and multilingual models is confounded by training-data quantity. Each monolingual model trains on a single speaker (1.7-11.8 hours), while every multilingual model trains on all speakers combined (about 23 hours) and, per Section 2, oversamples speakers so that each has roughly equal duration per epoch. This means the multilingual model sees several times more total audio and also sees the target speaker's own utterances many more times per epoch (e.g., Ojibwe NJ is duplicated several-fold to match Ojibwe JJ's duration). The improvements in MCD, LAS RMSE, and other metrics in Table 2 could therefore come from data quantity or from repeated exposure to the target speaker's own data, rather than from shared structure across typologically similar languages. To support the abstract's claim, the authors need a matched-data control (e.g., a monolingual model trained with the same amount of total audio and the same oversampling factor) and/or a control using typologically unrelated languages with the same total data. Without such controls, the statement in Section 5 that 'training on typologically similar languages does help alleviate the lack of data' is not established. The Limitations section does not acknowledge this confound.
  2. [Table 4 / Section 3.2] The architecture comparison is also confounded by hyperparameters. Table 4 shows that the Self-Attention model uses a learning rate of 1.00e-06 with no scheduler, while FNet, Mamba2, and Hydra all use a learning rate of 1.00e-04 with a cosine scheduler. The claim that attention-free architectures are 'highly competitive with self-attention' is therefore not a clean comparison of sequence-mixing layers; the observed differences could be due to learning-rate and scheduler choices. The authors should rerun the attention baseline with the same learning rate and scheduler used for the attention-free models, or provide a clear justification for why the different settings are necessary. At minimum, the learning-rate difference must be disclosed in the main text and factored into the interpretation.
  3. [Section 5 / Table 2] The paper reports no significance tests, confidence intervals, or multiple-seed variance, yet claims that multilingual models 'generally outperform monolingual models in all languages.' This is not uniformly supported by Table 2. For Ojibwe NJ, the monolingual Self-Attention model has a better F0 RMSE (80.511) than every multilingual variant (86.446 for Attention, 89.879 for Mamba2, 87.509 for Hydra, 97.015 for FNet), and the other metrics are mixed in sign. The claim that multilingual help is 'especially when data are scarce' is also not supported by the data pattern: the scarcest speaker (NJ) shows the most mixed results, while the largest improvement appears for Ojibwe JJ, the speaker with the most data. The authors should either provide statistical testing with multiple runs, or temper the claim to a descriptive statement about the specific runs reported.
  4. [Section 5 / Section 8 / Appendix E] The subjective evaluation is internally inconsistent. Section 8 states that the authors 'were not able to successfully conduct human MOS ratings,' which complicates interpretation, yet Section 5 states that 'the MOS ratings are largely consistent with the objective metrics.' Appendix E reports that only two participants completed the Ojibwe survey, with no ratings for Mi'kmaq and Maliseet, and that one Ojibwe rater assigned 5 to every Ojibwe NJ voice regardless of whether it was natural or synthetic. With one informative rater for one language, the statement that MOS ratings corroborate the objective results is unsupported. The paper should remove or substantially soften that sentence, or provide a more detailed per-rater analysis.
minor comments (7)
  1. [Appendix C] The caption says 'Tale 6' instead of 'Table 6'.
  2. [Throughout] The vocoder name is inconsistently typeset as 'V ocos' with a space; it should be 'Vocos' throughout the text and tables.
  3. [Section 8] The Limitations section uses 'Secondly' twice in the same list; the third item should begin with 'Thirdly.'
  4. [Table 2 and Table 6] The STOI column reports values around 0.03, which are implausible for the standard short-time objective intelligibility metric (which ranges from 0 to 1, with intelligible speech usually above 0.8). If STOI is intended, please correct the computation or rescale; if it is a different metric, please rename it and define it in the text.
  5. [Appendix A / Figure 1] The line immediately before Figure 1 contains a long series of unicode escape sequences ('/uni00000030/...'), which appears to be a font-encoding error in the PDF. The authors should ensure the appendix renders correctly.
  6. [Section 5] There is a typo: 'contemprary' should be 'contemporary.'
  7. [Table 4] The paper does not report the random seed or the number of training steps per model. Since epoch counts are fixed but dataset sizes differ, providing step counts and seeds would aid reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports empirical held-out measurements; no fitted parameter is renamed as a prediction and no load-bearing claim reduces to its own input by construction.

full rationale

The paper's central claim is that multilingual training on three related Algonquian languages improves TTS performance over monolingual models. This is an empirical comparison on held-out test utterances (Table 2), not a derivation from a fitted parameter or from an equation that assumes the result. The only self-citation is the recording protocol from Hammerly et al. 2023, used as an independent data-collection input; it does not justify the multilingual-transfer conclusion and is not load-bearing for the paper's main result. The comparison is confounded by data quantity: each monolingual model trains on one speaker's corpus while multilingual models train on all speakers' data, and oversampling is applied to balance durations. That is a genuine experimental-validity limitation, and the Limitations section does not acknowledge it, but it is not circularity under the criteria here: there is no exhibited reduction of an output to an input by definition, no fitted parameter renamed as a prediction, and no uniqueness theorem imported from the authors' own prior work. Therefore the circularity score is 0.

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

The central claims rest on the effectiveness of cross-lingual transfer, which is not directly tested against a data-quantity control, plus several standard-but-unvalidated assumptions about tokenization and evaluation metrics. No new theoretical entities are introduced.

free parameters (3)
  • Learning rate = 1e-6 (self-attention), 1e-4 (FNet/Mamba2/Hydra)
    Chosen separately per architecture; the 100x difference may affect the quality comparison.
  • Hidden dimension scaling = 3/4 reduction for Mamba2 and Hydra
    Shrunk to keep parameter counts near the attention baseline; this changes model capacity and could affect quality comparisons.
  • Oversampling target = Roughly equal duration per speaker
    Training samples duplicated to balance durations; this changes the data distribution and is not controlled against alternative ratios.
assumptions (4)
  • domain assumption Typological similarity (all three are Algonquian) is the mechanism enabling transfer
    The paper asserts this but never varies language relatedness; data quantity is a confound.
  • domain assumption Character-based tokenizer is adequate because the writing systems are phonetically transparent
    Invoked in Section 4; if orthographies are not fully transparent, TTS quality would degrade due to tokenization.
  • ad hoc to paper Oversampling by duplication does not harm training
    Duplicated samples can cause overfitting to repeated utterances; no analysis of this effect is provided.
  • domain assumption Objective metrics (F0 RMSE, MCD, PESQ, etc.) are valid proxies for synthesis quality in these languages
    Standard TTS metrics, but the paper itself notes MOS was mostly unobtainable, so the metrics are unvalidated against human judgments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Developing multilingual speech synthesis system for Ojibwe, Mi'kmaq, and Maliseet." pith.science (2026). https://pith.science/paper/6WK4SFSE

@misc{pith2026250202703,
  author       = {Pith},
  title        = {Pith review of: Developing multilingual speech synthesis system for Ojibwe, Mi'kmaq, and Maliseet},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6WK4SFSE}},
  note         = {Machine review of arXiv:2502.02703}
}
read the original abstract

We present lightweight flow matching multilingual text-to-speech (TTS) systems for Ojibwe, Mi'kmaq, and Maliseet, three Indigenous languages in North America. Our results show that training a multilingual TTS model on three typologically similar languages can improve the performance over monolingual models, especially when data are scarce. Attention-free architectures are highly competitive with self-attention architecture with higher memory efficiency. Our research not only advances technical development for the revitalization of low-resource languages but also highlights the cultural gap in human evaluation protocols, calling for a more community-centered approach to human evaluation.

Figures

Figures reproduced from arXiv: 2502.02703 by the authors.

Figure 1
Figure 1. Throughput comparison between different models and data types. Evaluations are all performed on a single A100 GPU with a batch size of 400. B Benchmarking efficiency Throughput We measured the throughput of each multilingual model with different data types (bfloat16, float16, and float32). Results are shown in [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. RTF comparison between different models and data types. Evaluations are all performed on a single A100 GPU. in every setting, followed by the Attention model, Mamba2 model, and Hydra model. D Vocoder comparison We compared three representative vocoders for waveform generation, namely, a time-domain vocoder HiFi-GAN (Kong et al., 2020), a frequency-domain vocoder Vocos (Siuzdak, 2024), and a diffusion-based vocoder F… view at source ↗
Figure 3
Figure 3. The MOS rating interface [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 18 canonical work pages

  1. [1]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Steven Bird. 2020. https://doi.org/10.18653/v1/2020.coling-main.313 Decolonising speech and language technology . In Proceedings of the 28th International Conference on Computational Linguistics, pages 3504--3519, Barcelona, Spain (Online). International Committee on Computational Linguistics

    doi:10.18653/v1/2020.coling-main.313 2020

  4. [4]

    William Chen, Takatomo Kano, Atsunori Ogawa, Marc Delcroix, and Shinji Watanabe. 2024. Train long and test long: Leveraging full document contexts in speech processing. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 13066--13070. IEEE

    work page 2024

  5. [5]

    Hyunjae Cho, Wonbin Jung, Junhyeok Lee, and Sang Hoon Woo. 2022. https://doi.org/10.21437/Interspeech.2022-46 Sane-tts: Stable and natural end-to-end multilingual text-to-speech . In Interspeech 2022, pages 1--5

    work page doi:10.21437/interspeech.2022-46 2022

  6. [6]

    Tri Dao and Albert Gu. 2024. Transformers are SSM s: Generalized models and efficient algorithms through structured state space duality. In International Conference on Machine Learning (ICML)

    2024

  7. [7]

    Christoph Draxler and Klaus J \"a nsch. 2004. Speechrecorder: A universal platform independent multi-channel audio recording software. In LREC

    work page 2004

  8. [8]

    Albert Gu and Tri Dao. 2023. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752

    arXiv 2023

Show all 31 references
  1. [9]

    Efficiently modeling long sequences with structured state spaces

    Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In International Conference on Learning Representations

  2. [10]

    Varun Gumma, Rishav Hada, Aditya Yadavalli, Pamir Gogoi, Ishani Mondal, Vivek Seshadri, and Kalika Bali. 2024. https://aclanthology.org/2024.computel-1.11 M un TTS : A text-to-speech system for M undari . In Proceedings of the Seventh Workshop on the Use of Computational Metho...

    2024

  3. [11]

    Christopher Hammerly, Sonja Foug \`e re, Giancarlo Sierra, Scott Parkhill, Harrison Porteous, and Chad Quinn. 2023. https://aclanthology.org/2023.computel-1.9 A text-to-speech synthesis system for border lakes O jibwe . In Proceedings of the Sixth Workshop on the Use of Comput...

    2023

  4. [12]

    Atticus Harrigan, Timothy Mills, and Antti Arppe. 2019. A preliminary plains cree speech synthesizer. In Proceedings of the Workshop on Computational Methods for Endangered Languages, volume 1

    2019

  5. [13]

    Sukjun Hwang, Aakash Lahoti, Tri Dao, and Albert Gu. 2024. Hydra: Bidirectional state space models through generalized matrix mixers. arXiv preprint arXiv:2407.09941

    2024 arXiv

  6. [14]

    Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and Fran c ois Fleuret. 2020. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156--5165. PMLR

    2020

  7. [15]

    Jungil Kong, Jaehyeon Kim, and Jaekyoung Bae. 2020. Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis. Advances in neural information processing systems, 33:17022--17033

    2020

  8. [16]

    James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, and Santiago Ontanon. 2022. https://doi.org/10.18653/v1/2022.naacl-main.319 FN et: Mixing tokens with F ourier transforms . In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational ...

    2022 doi

  9. [17]

    Xiang Li, FanBu FanBu, Ambuj Mehrish, Yingting Li, Jiale Han, Bo Cheng, and Soujanya Poria. 2024. https://doi.org/10.18653/v1/2024.findings-naacl.240 CM - TTS : Enhancing real time text-to-speech synthesis efficiency through weighted samplers and consistency models . In Findin...

    2024 doi

  10. [18]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. 2023. https://openreview.net/forum?id=PqvMRDCJT9t Flow matching for generative modeling . In The Eleventh International Conference on Learning Representations

    2023

  11. [19]

    Yuanjun Lv, Hai Li, Ying Yan, Junhui Liu, Danming Xie, and Lei Xie. 2024. https://arxiv.org/abs/2406.08196 Freev: Free lunch for vocoders through pseudo inversed mel filter . Preprint, arXiv:2406.08196

    2024 arXiv

  12. [20]

    Teresa L McCarty. 2018. Community-based language planning: Perspectives from indigenous language revitalization. In The Routledge handbook of language revitalization, pages 22--35. Routledge

    2018

  13. [21]

    Shivam Mehta, Ruibo Tu, Jonas Beskow, \'E va Sz \'e kely, and Gustav Eje Henter. 2024. Matcha-tts: A fast tts architecture with conditional flow matching. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 11341--11345. IEEE

    2024

  14. [22]

    Koichi Miyazaki, Yoshiki Masuyama, and Masato Murata. 2024. https://doi.org/10.21437/Interspeech.2024-994 Exploring the capability of mamba in speech applications . In Interspeech 2024, pages 237--241

    2024 doi

  15. [23]

    Tan Dat Nguyen, Ji-Hoon Kim, Youngjoon Jang, Jaehun Kim, and Joon Son Chung. 2024. Fregrad: Lightweight and fast frequency-aware diffusion vocoder. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 10736--10740. IEEE

    2024

  16. [24]

    Kari Noe and Nurit Kirshenbaum. 2024. Where generalized equitable design practice meet specific indigenous communities. In Proceedings of the CHI Conference on Human Factors in Computing Systems, pages 1--8

    2024

  17. [25]

    Richard T Oster, Angela Grier, Rick Lightning, Maria J Mayan, and Ellen L Toth. 2014. Cultural continuity, traditional indigenous language, and diabetes in alberta first nations: a mixed methods study. International journal for equity in health, 13:1--11

    2014

  18. [26]

    Aidan Pine, Erica Cooper, David Guzmán, Eric Joanis, Anna Kazantseva, Ross Krekoski, Roland Kuhn, Samuel Larkin, Patrick Littell, Delaney Lothian, Akwiratékha Martin, Koren Richmond, Marc Tessier, Cassia Valentini-Botinhao, Dan Wells, and Junichi Yamagishi. 2024. Speech genera...

    2024

  19. [27]

    Aidan Pine, Dan Wells, Nathan Brinklow, Patrick Littell, and Korin Richmond. 2022. https://doi.org/10.18653/v1/2022.acl-long.507 Requirements and motivations of low-resource speech synthesis for language revitalization . In Proceedings of the 60th Annual Meeting of the Associa...

    2022 doi

  20. [28]

    Henry Robertson. 2023. 2021 C ensus of population: Aboriginal peoples. G overnment of C anada

    2023

  21. [29]

    Hubert Siuzdak. 2024. https://openreview.net/forum?id=vY9nzQmQBw Vocos: Closing the gap between time-domain and fourier-based neural vocoders for high-quality audio synthesis . In The Twelfth International Conference on Learning Representations

    2024

  22. [30]

    Jeremy Zehr and Florian Schwarz. 2018. Penncontroller for internet based experiments (ibex)

    2018

  23. [31]

    Xiangyu Zhang, Qiquan Zhang, Hexin Liu, Tianyi Xiao, Xinyuan Qian, Beena Ahmed, Eliathamby Ambikairajah, Haizhou Li, and Julien Epps. 2024. Mamba in speech: Towards an alternative to self-attention. arXiv preprint arXiv:2405.12609

    2024 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.