Pith. sign in

REVIEW 4 major objections 5 minor 34 references

Optimizing Multilingual Text-To-Speech with Accents & Emotions

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims a text-to-speech system that switches between Hindi and Indian English inside one utterance while preserving emotion, reporting accent word error down from 15.4% to 11.8%.

desk verdict A fine-tuning write-up whose central code-switching mechanism is named but never specified, so the headline result is not checkable. read the letter →

arxiv 2506.16310 v1 pith:EHO4PVSY submitted 2025-06-19 cs.LG cs.HCcs.SDeess.AS

Pranav Pawar , Akshansh Dwivedi , Jenish Boricha , Himanshu Gohil , Aditya Dubey This is my paper
classification cs.LGcs.HCcs.SDeess.AS
keywords text-to-speechHindi-Englishcode-mixingaccentcontrolemotionmodelingtransliterationphonemealignmentvectorquantizationmultilingualTTS
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 attempts to show that a text-to-speech system can change accent and emotion inside a single sentence, so a speaker can begin in Indian English and slide into Hindi without pausing or sounding artificial. It builds on a prompt-conditioned text-to-speech model and adds a language-specific phoneme alignment layer, culture-sensitive emotion embedding layers, and a dynamic accent code switching module with residual vector quantization. The reported outcomes are a 23.7% relative reduction in accent word error (15.4% to 11.8%), 85.3% emotion recognition from native listeners, and a cultural correctness mean opinion score of 4.2 out of 5 from 200 users, all better than two existing multilingual emotional text-to-speech baselines. If these numbers hold, code-mixed speech synthesis becomes practical for South Asian education, accessibility tools, and localized virtual assistants.

What carries the argument

The load-bearing component is the dynamic accent code switching with residual vector quantization: a module intended to swap an accent code at language boundaries within a single utterance while residual vector quantization preserves the acoustic detail needed for continuity. Around it, the system stacks a language-specific phoneme alignment hybrid encoder-decoder and culture-sensitive emotion embedding layers trained on native speaker corpora. A frozen text encoder from a large language model turns natural-language descriptions such as 'calm, authoritative voice with a slight Indian accent' into conditioning vectors, and a neural vocoder renders the final audio. The paper's argument is that these components together separate accent from emotion, so both can be controlled independently and code-mixing can occur in real time.

What would settle it

Compare the full system with a version that leaves out the accent-switching component, using the paper's own mixed Hindi-English example; if the two versions score the same on word error and native-listener accent judgements, the claimed switching component is not what causes the improvement.

Watch

Extended reading notes

Core claim

The paper's central claim is that accent and emotion can be disentangled in a text-to-speech system, allowing real-time code-mixing: an utterance can start in Indian English and move into Hindi, as in 'Namaste, let's talk about <Hindi phrase>', with uninterrupted accent shifts and stable emotional tone. The authors report that their three-stage fine-tuning approach — Indian-accent English, then Hindi, then emotion — plus the dynamic accent code switching module yields a word error rate drop from 15.4% to 11.8% for accent accuracy, 85.3% emotion recognition by native listeners, and a cultural correctness MOS of 4.2/5 that they say surpasses existing multilingual systems. They also report strong control over gender and accent attributes, and audio quality close to ground truth on PESQ, STOI, and SISDR metrics.

Load-bearing premise

The whole headline feature depends on the assumption that the system really changes a speaker's accent when it moves from one language to another inside a sentence, since the paper describes the component that should do this but gives no technical details or test showing it works.

Editorial extensions

If this is right

  • Code-mixed Hindi-English speech can be generated in a single pass, so a user can request a sentence that starts in Indian English and moves into Hindi without retraining a model per speaker.
  • Accent and emotion are separable enough that improving accent accuracy (WER from 15.4% to 11.8%) does not degrade emotion recognition, which native listeners placed at 85.3%.
  • Natural-language prompts can control speaking rate, accent, and emotional tone together, enabling personalized output such as a calm, authoritative voice with a slight Indian accent.
  • The reported cultural correctness MOS of 4.2/5, significantly above existing multilingual systems, suggests that text-to-speech evaluation can treat cultural appropriateness as a distinct quality axis from audio fidelity.
  • The approach is intended to extend to additional Indic languages and regional dialects, with more training data and larger base models.

Reading between the lines

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

  • Inference: If the accent-code-switching module is real, the same three-stage recipe should transfer to other code-mixed language pairs, such as Spanish-English or Mandarin-English, by changing only the phoneme alignment and emotion embedding training data.
  • Inference: The paper does not report a per-boundary language-identification test, so a direct measure of accent shift timing at the code-switch point would be a sharper test than whole-utterance word error rate.
  • Inference: Because the Hindi training data is a five-hour male single-speaker corpus, the results likely understate performance for female voices, regional dialects, and spontaneous conversational prosody; scaling the data is the obvious next experiment.
  • Inference: The claimed real-time code-mixing could be verified by measuring latency and prosodic continuity at the boundary, neither of which the paper quantifies.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper describes a multilingual text-to-speech (TTS) system built on Parler-TTS, with added modules for language-specific phoneme alignment, culture-sensitive emotion embedding, and a 'dynamic accent code switching with residual vector quantization' mechanism, targeting Hindi and Indian English. The authors report quantitative improvements (23.7% relative WER reduction from 15.4% to 11.8%, 85.3% emotion recognition accuracy, MOS 4.2/5) and claim superiority over METTS and VECL-TTS baselines, with the headline feature being real-time mid-utterance code-mixed speech with uninterrupted accent shifts. The paper focuses on fine-tuning procedures and qualitative acoustic analyses, but does not provide an algorithmic description of the code-switching module or numeric comparisons against the stated baselines.

Significance. If fully substantiated, the ability to switch between Indian English and Hindi accents within a single utterance while maintaining emotional consistency would be a meaningful advance for Indic-language TTS and code-mixed speech synthesis. The paper targets a genuine gap in culturally aware multilingual TTS and includes native-listener evaluations as well as acoustic feature analyses. However, the significance is heavily tempered by the absence of a reproducible architecture specification and the lack of explicit baseline outcome data; the central claims are not independently verifiable from the present manuscript. The paper also ships no code, checkpoints, or test sets, so the results cannot be reproduced as described.

major comments (4)
  1. [§3.3–§3.5] The central novel module, 'dynamic accent code switching with residual vector quantization,' is never described algorithmically. Sections 3.3–3.5 present standard fine-tuning details (learning rates, batch sizes, loss functions) for Indian-accent English, Hindi, and emotion conditioning, but no equations, layer definitions, training objective, or inference-time switching logic are given for the code-switching mechanism. Since the abstract and §1 identify this module as the basis for the claimed 'uninterrupted accent shifts' and real-time code-mixing, its absence makes the primary contribution irreproducible and unfalsifiable.
  2. [§4] The abstract claims the proposed system is 'surpassing METTS and VECL-TTS baselines,' but Section 4 provides no numeric results for these baselines. There is no comparison table, no baseline WER/emotion/MOS values, and no description of how METTS and VECL-TTS runs were configured or evaluated. Without such data, the superiority claim cannot be quantitatively checked.
  3. [§4, Figure 8] The cross-lingual WER comparison shown in Figure 8 is presented without numeric labels, error bars, or sample sizes, and no statistical test is reported. The abstract's 'p < 0.01' for the MOS comparison is likewise unsupported by any test statistic or methodology description. As presented, the reader cannot determine whether the observed differences between the Hindi-English and English-Hindi configurations are reliable.
  4. [§3.1.6 and §4] The emotion categories used for training data annotation (e.g., 'sad,' 'happy,' 'laughing') appear to be the same categories tested in the emotion recognition evaluation reporting 85.3% accuracy from native listeners. This introduces circularity: the model is evaluated on its ability to reproduce the label distribution on which it was trained. The paper does not state whether the evaluation set was disjoint from the training data, how listeners were instructed, or whether the listeners were aware of the intended emotion labels.
minor comments (5)
  1. [Abstract and §4] The abstract states '23.7% improvement in accent accuracy (Word Error Rate reduction from 15.4% to 11.8%),' which is internally consistent as a relative WER reduction, but the phrase 'improvement in accent accuracy' is ambiguous for an error metric; consider stating explicitly that lower WER is better and that this is a relative reduction.
  2. [§1] The historical narrative in the Introduction (e.g., 'Al-Kindi’s invention of frequency analysis during the Arabic Crusade' and a 'Takumi' TTS system) is confusing and not supported by the cited references; this passage should be rewritten to accurately describe transliteration and its role in multilingual TTS.
  3. [§3.3] The text reports 'a learning rate of 1−4,' which is likely a typo for 1e-4 or 10^-4; numeric notations are inconsistent across the paper, mixing forms such as '5 × 10−5' and '8 × 10-5.' Please standardize.
  4. [Figures 4 and 7] Figures 4 and 7 appear to present identical plots of Mean Spectral Centroid, MFCC standard deviation, Zero-Crossing Rate, Energy, and Duration for the same four setups, and the text in §4 refers to 'Figure 5' when describing spectrograms that are captioned as Figure 7. The figure numbering and duplicated content need to be corrected.
  5. [§4] The statement that the model 'surpassed the Audiobox system' introduces a baseline that is not otherwise described or compared in any table; this term should either be defined with a reference or removed for clarity.

Circularity Check

2 steps flagged · score 6.0 of 10

The headline code-switching mechanism is a named module over standard VQ components, and the emotion model's training loss is reused as its own evaluation.

  1. renaming known result [Abstract; Section 3.2.3; Section 5]
    "as well as incorporating a dynamic accent code switching with residual vector quantization. ... For discrete modeling, a VQ-V AE is first trained to quantize the acoustic features. The acoustic model subsequently predicts these discrete tokens, typically with a Transformer-based architecture with discrete diffusion. ... It adjusts dynamically among accents, like- Indian English and Hindi language, in a single speech synthesis."

    Section 3.2.3 describes only the standard Parler/VQ-VAE discrete acoustic modeling recipe; no equation, layer, or inference rule specifies how an accent code is switched mid-utterance. The abstract's 'dynamic accent code switching with residual vector quantization' is therefore a label for a component whose behavior is never defined. The conclusion then credits that label with the system's central capability ('adjusts dynamically among accents'), so the result is supplied by the name rather than derived from the described architecture. This is a renaming of standard VQ-based acoustic modeling as the paper's novel contribution.

  2. fitted input called prediction [Section 3.5, Emotion-Based Model Fine-Tuning]
    "Cross-entropy loss function was employed to reduce the difference between predicted and real speech emotions. The ultimate model assessment had a loss of 3.27, which indicates its ability to produce emotionally rich speech outputs."

    The 3.27 loss is the value of the very objective used to fit the emotion model; minimizing it is the training procedure. Reporting that minimized loss as an 'ultimate model assessment' that 'indicates its ability' makes the evaluation measure the model's fit to its own training labels. A low value is partly forced by construction, so it does not independently establish emotional expressiveness.

full rationale

The paper has two concrete self-referential moves. First, the central 'dynamic accent code switching' is never instantiated: Section 3.2.3 presents standard VQ-VAE discrete acoustic modeling, and Sections 3.3-3.5 only document fine-tuning details. The capability that makes the system novel, uninterrupted mid-utterance accent shifts, is asserted by the module's name and then credited to the module in the conclusion, so the main claim reduces to definition rather than to a derivable mechanism. Second, the emotion model's final cross-entropy loss, the training objective, is used as evidence of emotional ability; that loss is minimized during fitting, so it is not an independent evaluation. The reported native-listener MOS and WER numbers could have served as independent evidence, but they are presented without protocols or comparison tables, so they do not break the circularity of the two moves above. There are no self-citations or imported uniqueness theorems; the circularity is primarily definitional/renaming and fitted-loss self-assessment.

Assumptions & free parameters 5 free parameters · 5 assumptions · 3 invented entities

The paper rests on pretrained models and hand-chosen labels. The central claims depend on emotion and accent tags chosen by the authors, a learned speaker embedding, and an unspecified RVQ codebook. No equations or ablations isolate the contribution of the new modules.

free parameters (5)
  • Emotion label set = 8 categories (whisper, enunciation, sad, default, laughing, confused, happy, emphasis)
    Chosen by hand for training; the model is conditioned on these labels, and the emotion recognition accuracy depends on this choice.
  • Speaker embedding table = not specified
    Learned identity embeddings for the single/multiple speakers used in fine-tuning; affects accent and voice consistency.
  • Accent embedding / code-switching codebook = not specified
    Residual vector quantization codebook for accent switching is claimed but its size and training are not described.
  • Loss weights = not specified
    The total loss sums mel-spectrogram reconstruction, duration, and pitch losses with unspecified weights; these affect the final model behavior.
  • Learning rates and batch sizes = 1e-4, 5e-5, 8e-5; batch 32/1
    Hand-chosen hyperparameters for the three fine-tuning stages.
assumptions (5)
  • domain assumption Pretrained Parler-TTS v0.1/mini-v1 provides a valid base for multilingual TTS fine-tuning.
    Sections 3.3-3.5 build on Parler-TTS without validating its suitability for Hindi and Indian English.
  • domain assumption Flan-T5 text encoder remains frozen and adequately represents the text descriptions.
    Section 3.1.9 states the text encoder is frozen and initialized from Flan-T5; this presumes its embeddings are sufficient.
  • domain assumption The dataspeech library's feature tags (speaking rate, SNR, reverberation, monotony) and the v01_bin_edges.json are appropriate for the new data.
    Sections 3.1.6-3.1.7 reuse these bins without re-estimating them for the new dataset.
  • domain assumption Word Error Rate measured by an ASR system is a valid proxy for accent accuracy.
    Section 4 reports WER for cross-lingual synthesis without specifying the ASR model or justifying this proxy.
  • domain assumption Native listener ratings reflect cultural correctness and emotion recognition.
    Sections 4 and the abstract rely on subjective MOS and emotion recognition scores without detailing the listening test protocol.
invented entities (3)
  • Dynamic accent code switching module with residual vector quantization
    purpose: To switch accents seamlessly in a single utterance (e.g., Hindi and English).
    The module is described only at a high level; no pseudocode, architecture diagram, or ablation shows it produces the claimed effect.
  • Language-specific phoneme alignment hybrid encoder-decoder
    purpose: To align phonemes for Hindi and English in the content encoder.
    No technical specification is given; the encoder-decoder is described generically in Section 3.2.
  • Culture-sensitive emotion embedding layers
    purpose: To embed emotions in a way sensitive to cultural context.
    No independent evidence or ablation is presented; the operation of these layers is not defined.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimizing Multilingual Text-To-Speech with Accents & Emotions." pith.science (2026). https://pith.science/paper/EHO4PVSY

@misc{pith2026250616310,
  author       = {Pith},
  title        = {Pith review of: Optimizing Multilingual Text-To-Speech with Accents & Emotions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EHO4PVSY}},
  note         = {Machine review of arXiv:2506.16310}
}
read the original abstract

State-of-the-art text-to-speech (TTS) systems realize high naturalness in monolingual environments, synthesizing speech with correct multilingual accents (especially for Indic languages) and context-relevant emotions still poses difficulty owing to cultural nuance discrepancies in current frameworks. This paper introduces a new TTS architecture integrating accent along with preserving transliteration with multi-scale emotion modelling, in particularly tuned for Hindi and Indian English accent. Our approach extends the Parler-TTS model by integrating A language-specific phoneme alignment hybrid encoder-decoder architecture, and culture-sensitive emotion embedding layers trained on native speaker corpora, as well as incorporating a dynamic accent code switching with residual vector quantization. Quantitative tests demonstrate 23.7% improvement in accent accuracy (Word Error Rate reduction from 15.4% to 11.8%) and 85.3% emotion recognition accuracy from native listeners, surpassing METTS and VECL-TTS baselines. The novelty of the system is that it can mix code in real time - generating statements such as "Namaste, let's talk about <Hindi phrase>" with uninterrupted accent shifts while preserving emotional consistency. Subjective evaluation with 200 users reported a mean opinion score (MOS) of 4.2/5 for cultural correctness, much better than existing multilingual systems (p<0.01). This research makes cross-lingual synthesis more feasible by showcasing scalable accent-emotion disentanglement, with direct application in South Asian EdTech and accessibility software.

Figures

Figures reproduced from arXiv: 2506.16310 by the authors.

Figure 1
Figure 1. Illustration of transliteration examples across English-Hindi. The columns represent source and target words. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Overview of the training phase of text-to-speech synthesis pipeline. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Training process for the emotion-based model, illustrating the flow from the base model, addition of accents, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: plots the audio features Mean Spectral Centroid, Standard Deviation of MFCCs, Mean Zero-Crossing Rate, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Relevance of our model in Emotions and Accents [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: illustrates the Mean Opinion Scores (MOS) for synthesized speech for three emotional categories: Happy, Sad, [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: plots audio features like Mean Spectral Centroid, Standard Deviation of MFCCs [34], Mean Zero-Crossing [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: plots the Word Error Rate (WER) for cross-lingual speech synthesis in Hindi-English and English-Hindi [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 25 canonical work pages

  1. [1]

    Transfer learning from speaker verification to multispeaker text-to-speech synthesis

    Ye Jia, Yu Zhang, Ron J Weiss, Quan Wang, Jonathan Shen, Fei Ren, Zhifeng Chen, Patrick Nguyen, Ruoming Pang, Ignacio Lopez Moreno, and Yonghui Wu. Transfer learning from speaker verification to multispeaker text-to-speech synthesis. In Advances in Neural Information Processing Systems , volume 31, pages 4485–4495. Curran Associates, Inc., 2018. arXiv:1806.04558

    arXiv 2018

  2. [2]

    Multilingual Text-to-Speech Synthesis for Turkic Languages Using Transliteration

    Rustem Yeshpanov, Saida Mussakhojayeva, and Yerbolat Khassanov. Multilingual text-to-speech synthesis for turkic languages using transliteration. In Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH, pages 5521–5525, 2023. arXiv:2305.15749

    work page Pith review arXiv 2023

  3. [3]

    Accented text-to-speech synthesis with limited data

    Xuehao Zhou, Mingyang Zhang, Yi Zhou, Zhizheng Wu, and Haizhou Li. Accented text-to-speech synthesis with limited data. IEEE/ACM Transactions on Audio, Speech, and Language Processing , 32:1699–1711, 2024. 10

    work page 2024

  4. [4]

    Multi-scale accent modeling with disen- tangling for multi-speaker multi-accent tts synthesis

    Xuehao Zhou, Mingyang Zhang, Yi Zhou, Zhizheng Wu, and Haizhou Li. Multi-scale accent modeling with disen- tangling for multi-speaker multi-accent tts synthesis. arXiv preprint arXiv:2406.10844, 2024. arXiv:2406.10844

    arXiv 2024

  5. [5]

    Text to speech synthesis: A systematic review, deep learning based architecture and future research direction

    Fahima Khanam, Farha Akhter Munmun, Nadia Afrin Ritu, Aloke Kumar Saha, and Muhammad Firoz. Text to speech synthesis: A systematic review, deep learning based architecture and future research direction. Journal of Advances in Information Technology, 13(5):398–412, 2022

    work page 2022

  6. [6]

    Weiss, Heiga Zen, Yonghui Wu, Zhifeng Chen, R

    Yu Zhang, Ron J. Weiss, Heiga Zen, Yonghui Wu, Zhifeng Chen, R. J. Skerry-Ryan, Ye Jia, Andrew Rosenberg, and Bhuvana Ramabhadran. Learning to speak fluently in a foreign language: Multilingual speech synthesis and cross-language voice cloning. arXiv preprint arXiv:1907.04448, 2019

    arXiv 1907

  7. [7]

    Explicit Intensity Control for Accented Text-to-speech

    Rui Liu, Haolin Zuo, De Hu, Guanglai Gao, and Haizhou Li. Explicit intensity control for accented text-to-speech. arXiv preprint arXiv:2210.15364, 2022

    work page Pith review arXiv 2022

  8. [8]

    Metts: Multilingual emotional text-to-speech by crossspeaker and cross-lingual emotion transfer

    Xinfa Zhu, Yi Lei, Tao Li, Yongmao Zhang, Hongbin Zhou, Heng Lu, and Lei Xie. Metts: Multilingual emotional text-to-speech by crossspeaker and cross-lingual emotion transfer. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 32:1506–1518, 2024

    work page 2024

Show all 34 references
  1. [9]

    Vecl-tts: voice identity and emotional style controllable cross-lingual text-to-speech

    Ashishkumar Gudmalwar, Nirmesh Shah, Sai Akarsh, Pankaj Wasnik, and Rajiv Ratn Shah. Vecl-tts: voice identity and emotional style controllable cross-lingual text-to-speech. arXiv preprint arXiv:2406.08076, 2024

    2024 arXiv

  2. [10]

    Ed-tts: Multi-scale emotion modeling using cross-domain emotion diarization for emotional speech synthesis

    Haobin Tang, Xulong Zhang, Ning Cheng, Jing Xiao, and Jianzong Wang. Ed-tts: Multi-scale emotion modeling using cross-domain emotion diarization for emotional speech synthesis. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP...

    2024

  3. [11]

    Advanced emotion and multi-speaker recognition with multilingual voicecloning in cross-cultural communication

    Jayapratha N, Vijaysurya M, Lingeshwaran G, Vema Naga Karish Gupta, and Shivaprasanna. Advanced emotion and multi-speaker recognition with multilingual voicecloning in cross-cultural communication. International Journal of Innovative Science and Research Technology (IJISRT) , ...

    2024

  4. [12]

    Exploring cross-linguistic speech perception in hindi, english, and romance-language through temporal dynamics of neural activity, 2024

    Yuga Kothari. Exploring cross-linguistic speech perception in hindi, english, and romance-language through temporal dynamics of neural activity, 2024

    2024

  5. [13]

    Diclet-tts: Diffusion model based crosslingual emotion transfer for text-to-speech—a study between english and mandarin

    Tao Li, Chenxu Hu, Jian Cong, Xinfa Zhu, Jingbei Li, Qiao Tian, Yuping Wang, and Lei Xie. Diclet-tts: Diffusion model based crosslingual emotion transfer for text-to-speech—a study between english and mandarin. IEEE/ACM Transactions on Audio, Speech, and Language Processing , ...

    2023

  6. [14]

    Diffsound: Discrete diffusion model for text-to-sound generation

    Dongchao Yang, Jianwei Yu, Helin Wang, Wen Wang, Chao Weng, Yuexian Zou, and Dong Yu. Diffsound: Discrete diffusion model for text-to-sound generation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:1720–1733, 2023

    2023

  7. [15]

    Spark- tts: An efficient llm-based text-to-speech model with single-stream decoupled speech tokens

    Xinsheng Wang, Mingqi Jiang, Ziyang Ma, Ziyu Zhang, Songxiang Liu, Linqin Li, Zheng Liang, et al. Spark- tts: An efficient llm-based text-to-speech model with single-stream decoupled speech tokens. arXiv preprint arXiv:2503.01710, 2025

    2025 arXiv

  8. [16]

    Nimisha Srivastava, Rudrabha Mukhopadhyay, and C. V . Jawahar. Indicspeech: Text-to-speech corpus for indian languages. In Proceedings of the Twelfth Language Resources and Evaluation Conference , pages 6417–6422, 2020

    2020

  9. [17]

    Natural language guidance of high-fidelity text-to-speech with synthetic annotations

    Dan Lyth and Simon King. Natural language guidance of high-fidelity text-to-speech with synthetic annotations. arXiv preprint arXiv:2402.01912, 2024

    2024 arXiv

  10. [18]

    Gemma 2: Improving open language models at a practical size

    Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024

    2024 arXiv

  11. [19]

    Scaling instruction-finetuned language models

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, et al. Scaling instruction-finetuned language models. Journal of Machine Learning Research, 25(70):1–53, 2024

    2024

  12. [20]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems , volume 30, 2017

    2017

  13. [21]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human langu...

    2019

  14. [22]

    Boosting fast and high-quality speech synthesis with linear diffusion

    Haogeng Liu, Tao Wang, Jie Cao, Ran He, and Jianhua Tao. Boosting fast and high-quality speech synthesis with linear diffusion. arXiv preprint arXiv:2306.05708, 2023

    2023 arXiv

  15. [23]

    Emovoice: Llm-based emotional textto-speech model with freestyle text prompting

    Guanrou Yang, Chen Yang, Qian Chen, Ziyang Ma, Wenxi Chen, Wen Wang, Tianrui Wang, et al. Emovoice: Llm-based emotional textto-speech model with freestyle text prompting. arXiv preprint arXiv:2504.12867, 2025. 11

    2025 arXiv

  16. [24]

    Vqalattent: a transparent speech generation pipeline based on transformer-learned vq-vae latent space

    Armani Rodriguez and Silvija Kokalj-Filipovic. Vqalattent: a transparent speech generation pipeline based on transformer-learned vq-vae latent space. arXiv preprint arXiv:2411.14642, 2024

    2024 arXiv

  17. [25]

    Indicvoices-r: Unlocking a massive multilingual multi- speaker speech corpus for scaling indian tts.Advances in Neural Information Processing Systems, 37:68161–68182, 2024

    Ashwin Sankar, Srija Anand, Praveen Varadhan, Sherry Thomas, Mehak Singal, Shridhar Kumar, Deovrat Mehendale, Aditi Krishana, Giri Raju, and Mitesh Khapra. Indicvoices-r: Unlocking a massive multilingual multi- speaker speech corpus for scaling indian tts.Advances in Neural In...

    2024

  18. [26]

    Detection of ai synthesized hindi speech

    Karan Bhatia, Ansh Agrawal, Priyanka Singh, and Arun Kumar Singh. Detection of ai synthesized hindi speech. arXiv preprint arXiv:2203.03706, 2022

    2022 arXiv

  19. [27]

    emotion2vec: Self-supervised pre-training for speech emotion representation

    Ziyang Ma, Zhisheng Zheng, Jiaxin Ye, Jinchao Li, Zhifu Gao, Shiliang Zhang, and Xie Chen. emotion2vec: Self-supervised pre-training for speech emotion representation. arXiv preprint arXiv:2312.15185, 2023

    2023 arXiv

  20. [28]

    Falk, Vijay Parsa, Joao F

    Tiago H. Falk, Vijay Parsa, Joao F. Santos, Kathryn Arehart, Oldooz Hazrati, Rainer Huber, James M. Kates, and Susan Scollie. Objective quality and intelligibility prediction for users of assistive listening devices: Advantages and limitations of existing tools. IEEE signal pr...

    2015

  21. [29]

    Timothy Bunnell, Ying Dou, Prasanna Kumar Muthukumar, Daniel Perry, Kishore Prahallad, Callie Vaughn, Alan W

    Stefan Steidl, Tim Polzehl, H. Timothy Bunnell, Ying Dou, Prasanna Kumar Muthukumar, Daniel Perry, Kishore Prahallad, Callie Vaughn, Alan W. Black, and Florian Metze. Emotion identification for evaluation of synthesized emotional speech. In Proc. speech prosody, volume 2012, p...

    2012

  22. [30]

    Stuck in the mos pit: A critical analysis of mos test methodology in tts evaluation

    Ambika Kirkland, Shivam Mehta, Harm Lameris, Gustav Eje Henter, Eva Székely, and Joakim Gustafson. Stuck in the mos pit: A critical analysis of mos test methodology in tts evaluation. In 12th Speech Synthesis Workshop (SSW), 2023

    2023

  23. [31]

    The limits of the mean opinion score for speech synthesis evaluation

    Sébastien Le Maguer, Simon King, and Naomi Harte. The limits of the mean opinion score for speech synthesis evaluation. Computer Speech & Language, 84:101577, 2024

    2024

  24. [32]

    Rix, John G

    Antony W. Rix, John G. Beerends, Michael P. Hollier, and Andries P. Hekstra. Perceptual evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and codecs. In 2001 IEEE international conference on acoustics, speech, and signal proce...

    2001

  25. [33]

    Identification and classification of tts intelligibility errors using asr: A method for automatic evaluation of speech intelligibility, 2023

    Erik Henriksson. Identification and classification of tts intelligibility errors using asr: A method for automatic evaluation of speech intelligibility, 2023

    2023

  26. [34]

    Tracey, D

    B. Tracey, D. V olfson, J. Glass, et al. Towards interpretable speech biomarkers: exploring mfccs. Sci Rep , 13:22787, 2023. 12

    2023

Pith tools

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