Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

LARoPE, a length-normalized rotary position embedding, biases cross-attention toward a diagonal text-speech alignment, lowering word error rate to 2.25% and staying stable to 30 seconds.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 17:05 UTC pith:WBPR5L5G

load-bearing objection LARoPE is a simple, promising TTS trick, but the paper's own math doesn't support its diagonal-bias claim—the empirical results may still hold, but the theory as written is wrong. the 3 major comments →

arxiv 2509.11084 v1 pith:WBPR5L5G submitted 2025-09-14 eess.AS cs.AIcs.CLcs.SD

Length-Aware Rotary Position Embedding for Text-Speech Alignment

Hyeongju Kim , Juheon Lee , Jinhyeok Yang , Jacob Morton This is my paper
classification eess.AS cs.AIcs.CLcs.SD
keywords text-to-speechtext-speech alignmentrotary position embeddingcross-attentionlength normalizationzero-shot TTSword error ratepositional encoding
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that a small change to rotary position embeddings—normalizing positions by sequence length—makes cross-attention in text-to-speech models learn better text-speech alignment. The method, called LARoPE, replaces absolute position offsets with length-normalized relative distances, which creates a diagonal bias in attention that matches the monotonic structure of speech. The authors show this leads to faster training convergence, fewer pronunciation errors, better speaker similarity, and stable long-form synthesis up to 30 seconds. This matters because attention-based TTS systems lack explicit alignment supervision, and alignment errors cause common failures like repetitions, omissions, and mispronunciations.

Core claim

The central claim is that RoPE, which encodes relative distance as the difference of absolute indices (m−n), is suboptimal for cross-attention because query and key sequences have different lengths. LARoPE instead rotates each subvector by an angle gamma*(p/L)*theta_j, where L is the sequence length and p the position. In cross-attention the inner product then depends on gamma*(m/Lq − n/Lk)*theta_j, meaning relative position is measured in normalized coordinates. This keeps the relative upper bound diagonal even when query and key lengths differ, and that diagonal conforms to the natural monotonic mapping between text and speech. Empirically, LARoPE consistently outperforms RoPE across word

What carries the argument

The length-aware rotation operator R′_theta(x, p, L), which rotates each 2D subvector by the angle gamma*(p/L)*theta_j instead of p*theta_j. This normalization turns the cross-attention relative term into gamma*(m/Lq − n/Lk)*theta_j, preserving a diagonal structure in the relative upper bound regardless of length mismatch. The diagonal bias steers attention toward proportional text-speech alignment and accelerates alignment learning without auxiliary losses or alignment tools.

Load-bearing premise

The method assumes the correct text-speech alignment is a straight diagonal in normalized coordinates—that speech frame m aligns with text token n approximately as m/Lq ≈ n/Lk—which can fail for speech with pauses, silence, or uneven speaking rates.

What would settle it

Train a TTS system with LARoPE on utterances with forced-aligner ground truth, then compare attention peaks to the aligner on prosodically unusual speech (long pauses, inserted breaths, very slow or fast speech). If LARoPE systematically shifts attention away from the aligner positions on those cases, the diagonal prior is too rigid.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • LARoPE reduces word error rate on both short and long utterances; on 10–30 second utterances it cuts WER from 4.98% to 2.16% when batch expansion is enabled.
  • It converges faster during training: with batch expansion disabled, CER at iteration 200k drops from 2.00% to 1.23%.
  • It stays robust when utterance duration is scaled from 0.7x to 1.4x, consistently beating RoPE in WER across all rates.
  • It achieves a state-of-the-art WER of 2.25% on the short-utterance zero-shot benchmark while keeping the same parameter count and inference cost as the base architecture.
  • Attention map analysis shows sharper, more continuous attention patterns under LARoPE, especially in early inference steps.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The diagonal prior assumes text and speech advance proportionally; real utterances with pauses, silences, or strongly variable phone durations may break this linear assumption, so comparing attention against forced-aligner ground truth could reveal where the prior is too rigid.
  • The scaling hyperparameter gamma controls the strength of the diagonal bias; tuning it per dataset or making it adaptive could expose a trade-off between alignment robustness and flexibility.
  • Because the mechanism only requires monotonic cross-modal alignment, the same length-normalized positional trick could transfer to other sequence-to-sequence tasks with linear ordering, such as speech recognition or monotonic machine translation, though those tasks would need their own validation.
  • LARoPE could be combined with explicit alignment strategies like guided attention or monotonic alignment search rather than replacing them, potentially yielding a stronger prior than either approach alone.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes LARoPE, a modification of rotary position embedding (RoPE) for cross-attention text-to-speech. Instead of rotating query/key vectors by p*theta_j, LARoPE uses gamma*(p/L)*theta_j, so the relative position term in the attention inner product becomes gamma*(m/Lq - n/Lk)*theta_j. The authors argue that this length-normalized relative distance induces a diagonal bias in cross-attention score maps, which better matches the monotonic structure of text-speech alignment. Experiments within the SupertonicTTS framework compare RoPE and LARoPE under two expansion factors (K_e=1 and 4) on LibriSpeech test-clean subsets (tc-short and tc-long), reporting WER, SIM, and UTMOSv2, along with training curves, duration-scaling robustness, attention-map visualizations, and a comparison with E2 TTS, F5-TTS, DiTTo-TTS, and SupertonicTTS. The headline result is a state-of-the-art 2.25% WER on the tc-short zero-shot benchmark while keeping the same parameter count and inference cost as SupertonicTTS.

Significance. If the mechanism holds, LARoPE is a simple and cost-free drop-in replacement for RoPE in cross-attention TTS, with particularly large gains on long utterances and duration variations. The empirical study is broad: two expansion factors, two length subsets, five duration scaling factors, training curves, and attention-map analysis. The paper is also transparent about the method definition and uses independent metrics (WER, SIM, UTMOS) rather than fitting a derivation to the results. However, the paper's central theoretical argument, as printed, does not mathematically support the diagonal-bias mechanism, and the empirical claims would be strengthened by uncertainty quantification and ablations. With correction of the derivation and additional robustness checks, this would be a solid contribution; as it stands, the key conceptual claim is not substantiated by the equations provided.

major comments (3)
  1. [Section 3, Eq. (4), Eq. (7), Fig. 1] The definition S_j = Sum_{k=0}^j |e^{i(m-n)theta_k}| is identically j+1 for all m and n, because |e^{ix}| = 1. The same applies to the LARoPE terms |e^{i*gamma*(m/Lq - n/Lk)*theta_j}|. Consequently, the 'relative upper bound' plotted in Fig. 1 is constant in (m,n); the diagonal structure shown cannot be obtained from the printed equations. The natural intended quantity is presumably S_j = |Sum_{k=0}^j e^{i*gamma*(m/Lq - n/Lk)*theta_k}|, whose magnitude does concentrate near m/Lq = n/Lk. This is load-bearing because the diagonal-bias mechanism is the paper's central conceptual contribution. The empirical results may still stand, but the theoretical argument as written does not substantiate the headline claim and must be corrected or explicitly revised.
  2. [Section 3, Section 4.2] The diagonal prior m/Lq = n/Lk assumes a strictly proportional relationship between text and speech positions. Real speech contains pauses, silence, and highly variable phone and word durations, so a linear proportional mapping can be a poor prior. The paper provides no comparison against aligner-derived ground truth, no ablation of the scaling parameter gamma, and no test of alternative monotonic biases (e.g., a learned affine or monotonic function). Without such evidence, it is unclear whether LARoPE's diagonal bias is genuinely aligned with natural speech or whether it is beneficial only because the test domain is approximately proportional. I recommend evaluating attention maps against forced alignments and ablating gamma and alternative monotonic parameterizations.
  3. [Section 4.2, Section 4.3, Tables 1-2] The central empirical claim is that LARoPE 'consistently outperforms' RoPE in WER and in duration-scaling robustness. However, WER results in Tables 1 and 2 are reported as point estimates with no confidence intervals or significance tests, and the paper describes only one training setup and seed. The durability of the gains, especially in Table 2 where some differences are small (e.g., 2.61 to 2.48 at 0.85d), cannot be assessed from the printed numbers. Additionally, gamma=10 is not ablated, leaving the sensitivity of the method to its main hyperparameter unknown. Adding multiple seeds or bootstrap confidence intervals for WER, and an ablation over gamma, would materially strengthen the claims.
minor comments (4)
  1. [Section 3, after Eq. (7)] The text says each term S_j is given by Sum_{k=0}^j |e^{i*gamma*(m/Lq - n/Lk)*theta_j}|, but the index of theta inside the sum over k should be theta_k, not theta_j.
  2. [Table 3] The DiTTo-TTS result is obtained on a 2.2-hour test-clean subset, whereas other baselines and the proposed model are evaluated on the tc-short set. This makes the comparison not strictly apples-to-apples. Please clarify or re-evaluate DiTTo-TTS on the same subset.
  3. [Section 4.1] The paper does not report whether the training data or the evaluation protocol includes any filtering of silence or pause-heavy utterances. Since the method's diagonal prior is most vulnerable to such cases, a brief description of data preprocessing and any silence handling would be useful.
  4. [General] No code, model checkpoints, or audio samples are provided. For a method whose primary claims are improved intelligibility and alignment, releasing samples or code would substantially aid reproducibility and verification.

Circularity Check

0 steps flagged

No significant circularity; the printed upper-bound argument is internally inconsistent but that is a correctness issue, not a circular derivation.

full rationale

LARoPE is defined directly by Equation (6) as a length-normalized rotation angle, and the attention form in Equation (7) follows algebraically. The scaling hyperparameter γ is fixed at 10, not fitted to the evaluation data, and the reported WER, SIM, CER, and UTMOS values are measured with external ASR, speaker-embedding, and quality models against independent test sets. The central claim—that LARoPE improves alignment and synthesis quality—is therefore empirically validated rather than derived from its own inputs. The only author-overlapping citation, SupertonicTTS [10], is used as the experimental framework and baseline, not as evidence for the method's mechanism, so self-citation is not load-bearing. One non-circular correctness flaw should be flagged: in Section 3, S_j is defined as a sum of absolute values of unit-modulus complex exponentials, so S_j = j+1 and the 'relative upper bound' is independent of m, n, L_q, and L_k; the claimed diagonal structure in Fig. 1 does not follow from the printed equations. This undermines the theoretical illustration but does not constitute circularity, because the empirical benchmark results are externally grounded and not equivalent to the definition of LARoPE.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The central method introduces one fitted hyperparameter (gamma), and rests on two domain assumptions: proportional text-speech alignment and the usefulness of a diagonal attention prior. No new physical entities are postulated.

free parameters (1)
  • gamma = 10
    Scaling hyperparameter in Eq. (6) that controls the length-normalized rotation frequency. It is set by hand to 10 with no ablation or sensitivity analysis. All LARoPE attention properties and the reported results depend on this value.
axioms (4)
  • standard math RoPE inner product equals Re(sum_j q_j k_j* e^{i(m-n)theta_j})
    Used in Section 2, Eq. (3), as the basis for relative position encoding in self-attention; standard property of the rotation operator.
  • standard math The geometric-series sum of unit-magnitude exponentials bounds the attention score, and this bound is maximized on the diagonal m/Lq = n/Lk
    Used in Section 3 and Fig. 1 to argue that LARoPE preserves a diagonal structure. The argument compares an upper bound, not the actual content-dependent attention scores.
  • domain assumption Text-speech alignment is approximately proportional: speech frame m aligns to text token n such that m/Lq is close to n/Lk
    Central modeling prior behind the length-normalized indices in Eq. (7). If real alignments involve pauses, nonlinear speaking rate, or leading/trailing silence, the proportional diagonal can be an incorrect bias.
  • domain assumption A diagonal-bias attention map improves alignment learning and final TTS intelligibility
    Adopted as the design goal in Sections 1 and 3, and supported by attention-map visualization in Section 4.4. It is not independently proven and no comparison against alternative monotonic biases is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 8002 in / 10508 out tokens · 121148 ms · 2026-08-04T17:05:22.038565+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Length-Aware Rotary Position Embedding for Text-Speech Alignment." pith.science (2026). https://pith.science/paper/WBPR5L5G

@misc{pith2026250911084,
  author       = {Pith},
  title        = {Pith review of: Length-Aware Rotary Position Embedding for Text-Speech Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WBPR5L5G}},
  note         = {Machine review of arXiv:2509.11084}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Many recent text-to-speech (TTS) systems are built on transformer architectures and employ cross-attention mechanisms for text-speech alignment. Within these systems, rotary position embedding (RoPE) is commonly used to encode positional information in text and speech representations. In this work, we introduce length-aware RoPE (LARoPE), a simple yet effective extension of RoPE that improves text-speech alignment. Unlike RoPE, which relies on absolute indices, LARoPE computes relative distances between query and key positions using length-normalized indices. Experimental results show that LARoPE consistently outperforms RoPE, offering faster loss convergence, more accurate text-speech alignment, and higher overall TTS quality. Furthermore, LARoPE demonstrates greater resilience to variations in utterance duration and maintains stable performance in extended speech generation up to 30 seconds, whereas RoPE suffers from notable degradation. Notably, our method achieves a state-of-the-art word error rate on a standard zero-shot TTS benchmark.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. RobustSpeechFlow: Learning Robust Text-to-Speech Trajectories via Augmentation-based Contrastive Flow Matching

    cs.SD 2026-05 unverdicted novelty 6.0

    RobustSpeechFlow improves TTS alignment robustness by extending contrastive flow matching with length-preserving repeat and skip latent augmentations, lowering WER from 1.44 to 1.38 on Seed-TTS-eval and CER on ZERO500.

  2. RobustSpeechFlow: Learning Robust Text-to-Speech Trajectories via Augmentation-based Contrastive Flow Matching

    cs.SD 2026-05 conditional novelty 6.0

    By training flow-matching TTS to avoid augmented repeat/skip latent trajectories, RobustSpeechFlow cuts Seed-TTS-eval WER from 1.44 to 1.38 and improves CER on a new multilingual benchmark.

Reference graph

Works this paper leans on

33 extracted references · 8 linked inside Pith · cited by 1 Pith paper

  1. [1]

    While some speech datasets provide phoneme-level timestamps, such detailed annota- tions are not available in most speech corpora [1, 2, 3, 4]

    INTRODUCTION Optimization of generative models for text-to-speech (TTS) requires accurate alignments between text and speech. While some speech datasets provide phoneme-level timestamps, such detailed annota- tions are not available in most speech corpora [1, 2, 3, 4]. To address unknown text-speech alignments, several techniques have been ex- plored in t...

  2. [2]

    PRELIMINARY RoPE [14] is a sophisticated approach to positional encoding that ef- fectively captures sequential information in transformer models. Un- like traditional positional embeddings which explicitly add position- dependent vectors to token representations, RoPE encodes positional information implicitly through a rotational transformation applied t...

    Pith/arXiv arXiv 2025

  3. [3]

    In this setup, speech features act as queries, while text embeddings serve as keys

    PROPOSED METHOD Contemporary TTS models are usually built on transformer blocks and employ cross-attention layers to condition speech features on textual representations [10, 11]. In this setup, speech features act as queries, while text embeddings serve as keys. To incorporate po- sitional information, RoPE is typically applied to both queries and keys b...

  4. [4]

    Pos. emb

    EXPERIMENTS 4.1. Experimental setup We evaluate the effectiveness of LARoPE within the SupertonicTTS framework [10]. SupertonicTTS introduces batch expansion to ac- celerate text-speech alignment. However, since our primary goal is to assess the impact of LARoPE on alignment itself, we also con- sider a baseline model without batch expansion (i.e., expans...

  5. [5]

    RELA TED WORK 5.1. Rotary position embedding RoPE encodes relative positions by rotating query and key vectors and has become the default positional encoding in transformer- based speech and language models [14, 10, 11]. Beyond the original formulation, several extensions such as position interpolation [22], YaRN [23], and LongRoPE [24] have been proposed...

  6. [6]

    As a length-aware extension of rotary position embed- dings, LARoPE normalizes positional indices by sequence length when computing relative distances between query and key vectors

    CONCLUSION We introduced LARoPE, a method designed to enhance text-speech alignment in TTS models that rely on cross-attention for text con- ditioning. As a length-aware extension of rotary position embed- dings, LARoPE normalizes positional indices by sequence length when computing relative distances between query and key vectors. This strategy induces a...

  7. [7]

    Timit acoustic phonetic continuous speech corpus,

    John S Garofolo, “Timit acoustic phonetic continuous speech corpus,”Linguistic Data Consortium, 1993, 1993

    1993

  8. [8]

    Libritts: A corpus derived from librispeech for text-to-speech,

    Heiga Zen, Viet Dang, Rob Clark, Yu Zhang, Ron J Weiss, Ye Jia, Zhifeng Chen, and Yonghui Wu, “Libritts: A corpus derived from librispeech for text-to-speech,”arXiv preprint arXiv:1904.02882, 2019

    Pith/arXiv arXiv 1904

  9. [9]

    Librispeech: An asr corpus based on public do- main audio books,

    Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur, “Librispeech: An asr corpus based on public do- main audio books,” in2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015, pp. 5206–5210

    2015

  10. [10]

    CSTR VCTK Corpus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92),

    Junichi Yamagishi, Christophe Veaux, and Kirsten MacDon- ald, “CSTR VCTK Corpus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92),” 2019

    2019

  11. [11]

    Diff-tts: A denoising dif- fusion model for text-to-speech,

    Myeonghun Jeong, Hyeongju Kim, Sung Jun Cheon, By- oung Jin Choi, and Nam Soo Kim, “Diff-tts: A denoising dif- fusion model for text-to-speech,” inInterspeech 2021, 2021, pp. 3605–3609

    2021

  12. [12]

    Glow-tts: A generative flow for text-to-speech via monotonic alignment search,

    Jaehyeon Kim, Sungwon Kim, Jungil Kong, and Sungroh Yoon, “Glow-tts: A generative flow for text-to-speech via monotonic alignment search,”Advances in Neural Informa- tion Processing Systems, vol. 33, pp. 8067–8077, 2020

    2020

  13. [13]

    DualSpeech: En- hancing Speaker-Fidelity and Text-Intelligibility Through Dual Classifier-Free Guidance,

    Jinhyeok Yang, Junhyeok Lee, Hyeong-Seok Choi, Seunghoon Ji, Hyeongju Kim, and Juheon Lee, “DualSpeech: En- hancing Speaker-Fidelity and Text-Intelligibility Through Dual Classifier-Free Guidance,” inInterspeech 2024, 2024, pp. 4423–4427

    2024

  14. [14]

    Neural codec language models are zero-shot text to speech synthesizers,

    Chengyi Wang, Sanyuan Chen, Yu Wu, Ziqiang Zhang, Long Zhou, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, et al., “Neural codec language models are zero-shot text to speech synthesizers,”arXiv preprint arXiv:2301.02111, 2023

    Pith/arXiv arXiv 2023

  15. [15]

    Speak, read and prompt: High-fidelity text-to-speech with minimal supervision,

    Eugene Kharitonov, Damien Vincent, Zal ´an Borsos, Rapha ¨el Marinier, Sertan Girgin, Olivier Pietquin, Matt Sharifi, Marco Tagliasacchi, and Neil Zeghidour, “Speak, read and prompt: High-fidelity text-to-speech with minimal supervision,”Trans- actions of the Association for Computational Linguistics, vol. 11, pp. 1703–1718, 2023

    2023

  16. [16]

    Supertonictts: Towards highly scalable and efficient text-to- speech system,

    Hyeongju Kim, Jinhyeok Yang, Yechan Yu, Seunghun Ji, Ja- cob Morton, Frederik Bous, Joon Byun, and Juheon Lee, “Supertonictts: Towards highly scalable and efficient text-to- speech system,”arXiv preprint arXiv:2503.23108, 2025

    arXiv 2025

  17. [17]

    DiTTo-TTS: Diffusion transformers for scalable text-to-speech without domain-specific factors,

    Keon Lee, Dong Won Kim, Jaehyeon Kim, Seungjun Chung, and Jaewoong Cho, “DiTTo-TTS: Diffusion transformers for scalable text-to-speech without domain-specific factors,” in The Thirteenth International Conference on Learning Repre- sentations, 2025

    2025

  18. [18]

    F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,

    Yushen Chen, Zhikang Niu, Ziyang Ma, Keqi Deng, Chunhui Wang, Jian Zhao, Kai Yu, and Xie Chen, “F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,”arXiv preprint arXiv:2410.06885, 2024

    Pith/arXiv arXiv 2024

  19. [19]

    E2 tts: Embarrassingly easy fully non-autoregressive zero-shot tts,

    Sefik Emre Eskimez, Xiaofei Wang, Manthan Thakker, Canrun Li, Chung-Hsien Tsai, Zhen Xiao, Hemin Yang, Zirun Zhu, Min Tang, Xu Tan, et al., “E2 tts: Embarrassingly easy fully non-autoregressive zero-shot tts,” in2024 IEEE Spoken Lan- guage Technology Workshop (SLT). IEEE, 2024, pp. 682–689

    2024

  20. [20]

    Roformer: Enhanced transformer with rotary position embedding,

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu, “Roformer: Enhanced transformer with rotary position embedding,”Neurocomputing, vol. 568, pp. 127063, 2024

    2024

  21. [21]

    The lj speech dataset,

    Keith Ito and Linda Johnson, “The lj speech dataset,”https: //keithito.com/LJ-Speech-Dataset/, 2017

    2017

  22. [22]

    Hi-Fi Multi-Speaker English TTS Dataset,

    Evelina Bakhturina, Vitaly Lavrukhin, Boris Ginsburg, and Yang Zhang, “Hi-Fi Multi-Speaker English TTS Dataset,” arXiv preprint arXiv:2104.01497, 2021

    Pith/arXiv arXiv 2021

  23. [23]

    Decoupled weight decay regularization,

    Ilya Loshchilov and Frank Hutter, “Decoupled weight decay regularization,” inInternational Conference on Learning Rep- resentations, 2019

    2019

  24. [24]

    Nemo inverse text normalization: From develop- ment to production,

    Yang Zhang, Evelina Bakhturina, Kyle Gorman, and Boris Ginsburg, “Nemo inverse text normalization: From develop- ment to production,”arXiv preprint arXiv:2104.05055, 2021

    Pith/arXiv arXiv 2021

  25. [25]

    Wavlm: Large-scale self-supervised pre-training for full stack speech processing,

    Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shu- jie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yosh- ioka, Xiong Xiao, et al., “Wavlm: Large-scale self-supervised pre-training for full stack speech processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505– 1518, 2022

    2022

  26. [26]

    The t05 system for the V oiceMOS Challenge 2024: Transfer learning from deep image classifier to natu- ralness MOS prediction of high-quality synthetic speech,

    Kaito Baba, Wataru Nakata, Yuki Saito, and Hiroshi Saruwatari, “The t05 system for the V oiceMOS Challenge 2024: Transfer learning from deep image classifier to natu- ralness MOS prediction of high-quality synthetic speech,” in IEEE Spoken Language Technology Workshop (SLT), 2024

    2024

  27. [27]

    Librispeech-pc: Benchmark for evaluation of punctuation and capitalization capabilities of end-to-end asr models,

    Aleksandr Meister, Matvei Novikov, Nikolay Karpov, Evelina Bakhturina, Vitaly Lavrukhin, and Boris Ginsburg, “Librispeech-pc: Benchmark for evaluation of punctuation and capitalization capabilities of end-to-end asr models,” in2023 IEEE automatic speech recognition and understanding work- shop (ASRU). IEEE, 2023, pp. 1–7

    2023

  28. [28]

    Extending context window of large lan- guage models via positional interpolation,

    Shouyuan Chen, Sherman Wong, Liangjian Chen, and Yuan- dong Tian, “Extending context window of large lan- guage models via positional interpolation,”arXiv preprint arXiv:2306.15595, 2023

    Pith/arXiv arXiv 2023

  29. [29]

    Yarn: Efficient context window extension of large lan- guage models,

    Bowen Peng, Jeffrey Quesnelle, Honglu Fan, and Enrico Ship- pole, “Yarn: Efficient context window extension of large lan- guage models,”arXiv preprint arXiv:2309.00071, 2023

    Pith/arXiv arXiv 2023

  30. [30]

    Lon- groPE: Extending LLM context window beyond 2 million to- kens,

    Yiran Ding, Li Lyna Zhang, Chengruidong Zhang, Yuanyuan Xu, Ning Shang, Jiahang Xu, Fan Yang, and Mao Yang, “Lon- groPE: Extending LLM context window beyond 2 million to- kens,” inForty-first International Conference on Machine Learning, 2024

    2024

  31. [31]

    Efficiently trainable text-to-speech system based on deep convolutional networks with guided attention,

    Hideyuki Tachibana, Katsuya Uenoyama, and Shunsuke Ai- hara, “Efficiently trainable text-to-speech system based on deep convolutional networks with guided attention,” in2018 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2018, pp. 4784–4788

    2018

  32. [32]

    Rad-tts: Parallel flow-based tts with robust alignment learning and diverse synthesis,

    Kevin J Shih, Rafael Valle, Rohan Badlani, Adrian Lancucki, Wei Ping, and Bryan Catanzaro, “Rad-tts: Parallel flow-based tts with robust alignment learning and diverse synthesis,” in ICML Workshop on Invertible Neural Networks, Normalizing Flows, and Explicit Likelihood Models, 2021

    2021

  33. [33]

    One tts alignment to rule them all,

    Rohan Badlani, Adrian Ła ´ncucki, Kevin J Shih, Rafael Valle, Wei Ping, and Bryan Catanzaro, “One tts alignment to rule them all,” inICASSP 2022-2022 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 6092–6096

    2022