Pith. sign in

REVIEW 3 major objections 5 minor 54 references

Kinship in Speech: Leveraging Linguistic Relatedness for Zero-Shot TTS in Indian Languages

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read By matching text-parsing rules and source-language choice to an unseen language's phonotactics, zero-shot TTS can produce natural speech for Sanskrit, Konkani dialects, Maithili, and Kurukh with no training data.

desk verdict A useful zero-shot TTS extension for five Indian languages, but the evaluation confounds parser-rule choice with source TTS identity. read the letter →

arxiv 2506.03884 v1 pith:KBSEHJNQ submitted 2025-06-04 cs.CL cs.CV

Utkarsh Pathak , Chandra Sai Krishna Gunda , Anusha Prakash , Keshav Agarwal , Hema A. Murthy This is my paper
classification cs.CLcs.CV
keywords zero-shotTTScommonlabelsetIndianlanguagesphonotacticsgrapheme-to-phonemelow-resourcespeechsynthesiscross-lingualtransferlanguagefamilies
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 aims to show that zero-shot text-to-speech is possible for Indian languages with almost no digital resources, by choosing a source language that is linguistically related to the target and adapting the text-to-phone pipeline to the target's own sound patterns. The key move is to stop parsing text by its written script and instead parse by the target language's phonotactics, using an expanded shared phone inventory (the common label set) and either Indo-Aryan or Dravidian grapheme-to-phoneme rules. On Sanskrit, two Konkani dialects, Maithili, and Kurukh, the right combination of source TTS and parsing rules produces intelligible, natural-sounding speech without any target-language training data, with best mean opinion scores of 4.12 (Sanskrit via Kannada) and 4.02 (Kurukh via Hindi). If this holds, TTS could be extended to many under-represented languages by exploiting linguistic kinship rather than collecting studio data.

What carries the argument

The central object is the common label set (CLS), a shared inventory of 72 phone labels that maps acoustically similar phones across Indian languages; the paper expands it by substituting sounds missing from the source scripts with the closest equivalents (e.g., uvular /q/ to velar /k/, voiced uvular fricative /ɣ/ to /g/, retroflex flap to dental flap). The second mechanism is the unified grapheme-to-phoneme parser, originally with separate Indo-Aryan and Dravidian rules, which is modified so the chosen rule family follows the target language's phonotactics—schwa retention for Sanskrit and Canara Konkani, schwa deletion for Maharashtrian Konkani, Maithili, and Kurukh—plus a few language-specific rules such as Sanskrit visarga assimilation and Maithili epenthesis. A suitable monolingual TTS system is then selected on the basis of which rule family was used and geographic/structural kinship, so the source voice's phone set and prosody align with the target.

What would settle it

A direct ablation would settle it: synthesise the same unseen-language text with the 'wrong' family rules (Hindi/IA for Sanskrit, Kannada/DR for Maharashtrian Konkani) while keeping the same TTS source and voice; if SUS word-error rates and MOS stay statistically unchanged, then phonetic kinship, not source selection, is doing the work.

Watch

Extended reading notes

Core claim

Zero-shot TTS can work for languages whose script and phonotactics belong to different families if the system is built around a shared phone representation augmented to cover the target language's missing sounds, and if text parsing rules are chosen to match the target language's phonotactics rather than its script. The paper demonstrates this by synthesising Sanskrit, Maharashtrian and Canara Konkani, Maithili, and Kurukh using monolingual Hindi, Kannada, Marathi, or Telugu TTS systems, with no adaptation data from the target languages. The results show that a Dravidian source (Kannada or Telugu) with Dravidian parsing rules suits Sanskrit because Sanskrit retains schwas, while Indo-Aryan source (Hindi) with IA rules suits Kurukh because Kurukh deletes schwas; dialect-level choice also matters, with Marathi better for Maharashtrian Konkani and Kannada better for Canara Konkani.

Load-bearing premise

The load-bearing premise is that the hand-assigned phone substitutions and the hand-selected parsing-rule family (IA or DR) for each target language really capture that language's phonotactics, because a wrong substitution or wrong rule family would make the speech unintelligible no matter which source TTS is used.

Editorial extensions

If this is right

  • Sanskrit, despite being written in Devanagari, is better synthesised by Dravidian-family systems (Kannada, Telugu) with Dravidian parsing rules than by Hindi, because Sanskrit lacks schwa deletion.
  • Dialect differences can be captured zero-shot: Maharashtrian Konkani scores higher with a Marathi system, while Canara Konkani scores higher with a Kannada system.
  • Kurukh, a Dravidian language written in Devanagari with IA-like schwa deletion, synthesises better from Hindi with IA rules than from Kannada.
  • The approach removes the requirement of target-language training data, reducing TTS development for a new language to choosing a relative source and adapting parsing rules.
  • The same language-centric procedure is claimed to extend to other low-resource languages and dialects beyond the Indian context.

Reading between the lines

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

  • An implication the paper leaves implicit is that the hand-crafted CLS substitutions and parsing-rule choices could be automated by measuring phonetic distance or using typological databases, making the pipeline scalable to hundreds of languages.
  • The paper's logic predicts a specific ordering for other script-phonotactics mismatched languages it mentions, such as Bodo (Sino-Tibetan, Devanagari script) or Kui/Kuvi (Dravidian, Odia script); a decisive extension would be to synthesise these and check whether the predicted source languages win.
  • Because the evaluations use text supplied to listeners, the reported WER may partly reflect a given-to-answer task; an extension without displaying the text would test whether intelligibility survives unaided listening.
  • The success of Kannada for Sanskrit suggests prosodic/rhythmic similarity across families may matter more than lexical closeness, which could guide source selection for other sprachbund areas.
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 / 5 minor

Summary. The paper proposes a zero-shot TTS pipeline for unseen Indian languages. It extends a common label set (CLS) with hand-specified phone substitutions, selects Indo-Aryan (IA) or Dravidian (DR) grapheme-to-phoneme parsing rules based on the target language's phonotactics rather than its script, and chooses a source monolingual TTS system from Hindi, Kannada, Marathi, or Telugu. The method is evaluated on Sanskrit, Maharashtrian Konkani, Canara Konkani, Maithili, and Kurukh using MOS, SUS WER, and MCD, with the claim that the combined parser-rule and source-system choices yield intelligible, natural speech without target-language training data.

Significance. If the central claim is established, the paper would provide a practical, low-cost route to TTS for under-resourced Indian languages and a compelling demonstration that typological and phonotactic knowledge can substitute for adaptation data. The study has genuine strengths: the parser-rule assignments and phone substitutions are specified a priori from documented linguistic facts rather than fitted to the evaluation outcomes; the evaluations use held-out text and native-speaker listeners; and audio samples are provided. However, the experiments do not cleanly separate the effect of the parser-rule modification from the effect of the source TTS system, and the quantitative evidence is too thin to support the causal attribution that the paper advertises.

major comments (3)
  1. [5.2, Tables 1-2] The experimental design conflates the parser-rule family with the source TTS identity for the comparisons that most directly support the paper's novelty claim. For Sanskrit, Sys 1 (Hindi, IA rules) is compared with Sys 2 (Kannada, DR rules) and Sys 3 (Telugu, DR rules); for Maharashtrian Konkani, Sys 1 (Marathi, IA) is compared with Sys 2 (Kannada, DR). In both cases the parser family and the source synthesiser change together, so the better MOS/WER could be caused by the intrinsic quality or prosody of the source system rather than by the proposed parsing. The Canara Konkani condition (Marathi, DR vs. Kannada, DR) and the Kurukh condition (Hindi, IA vs. Kannada, IA) hold the parser family fixed while changing the source, so they do not supply the missing control. A cell that holds the source TTS fixed and toggles only the IA/DR parser rules (e.g., Hindi with IA vs. Hindi with DR for Sanskrit, or Kannada with IA vs. Kannada with DR for Konkani) is needed to attribute the observed gains to the phonotactics-aware parsing.
  2. [5.3-5.5, Tables 1-3] The quantitative evidence does not yet support the relative claims. Evaluator counts are very small for Canara Konkani (6 listeners for MOS, 5 for SUS), and no confidence intervals or significance tests are reported for any language. More importantly, for Maharashtrian Konkani the MOS ordering (Marathi 3.64 vs. Kannada 2.54 in Table 1) is reversed by the MCD ordering (Kannada 8.28 vs. Marathi 10.74 in Table 3); the text attributes this to 'artefacts introduced by the Marathi system' but provides no analysis. Since the paper's central conclusion is that the linguistically chosen system is better, this contradiction and the absence of uncertainty quantification need to be resolved.
  3. [4.2-4.3] The CLS substitutions and parser-rule additions are hand-assigned and are not validated as a component. For instance, the mapping of /q/ to /k/ and of 'ड़' (/ɽ/) to /d/ is asserted to be the closest equivalence, but no check is reported for whether these choices erase phonemic distinctions in the target language. An ablation that varies the substitutions and rules while keeping the source TTS fixed, or a targeted error analysis on minimal pairs, would show that the specific linguistic choices, rather than generic cross-lingual transfer, drive intelligibility.
minor comments (5)
  1. [3] There is a typo: 'haveschwa' should be 'have schwa'.
  2. [5.3] 'quite environment' should be 'quiet environment'.
  3. [Abstract] The index term 'T erms' has an extra space.
  4. [4.2] The full CLS mapping and substitution details are delegated to an external URL; including the complete mapping in the paper or an appendix would improve reproducibility.
  5. [5.3] The ground-truth audio conditions differ across languages (studio, quiet room, Bible audio processed with VoiceFixer); this should be stated as a limitation when interpreting GT MOS scores and MCD comparisons across languages.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the zero-shot TTS results are genuine held-out evaluations, and the linguistic-kinship selection rules are stated before evaluation rather than fitted to the reported MOS/WER scores.

full rationale

The paper's central claim is that choosing a shared CLS phone set, family-specific G2P rules, and a source TTS based on linguistic kinship enables zero-shot synthesis for unseen Indian languages. Tracing the chain from Sections 4.2-4.4 to Tables 1-3 shows no load-bearing step that reduces to its own input. The CLS phone substitutions (/q/ to /k/, /ɣ/ to /g/, /ɽ/ to /d/) are hand-specified linguistic hypotheses, stated in Section 4.2 before any evaluation; they are not fitted to the MOS/WER/MCD columns. The parser-rule assignments (DR rules for Sanskrit and Canara Konkani, IA rules for Maharashtrian Konkani, Maithili, and Kurukh) are likewise justified in Sections 3 and 4.3 from independently documented properties (Sanskrit lacks schwa deletion; Kurukh has schwa deletion and IA borrowings; Konkani dialect differences), and the evaluation then tests these assignments on held-out native-speaker judgments. The reported scores are therefore genuine empirical observations, not quantities defined by construction. Self-citations to prior CLS and unified-parser work (e.g., [24], [36]) supply the shared representation, but those tools are described and modified in the paper, and the target-language choices are not justified by pointing to the current results. Two experimental limitations are worth noting, but they are not circularity: (1) Tables 1-3 vary parser-rule family and source synthesiser simultaneously, so the causal role of parser choice is not cleanly isolated; (2) for Maharashtrian Konkani, MOS favors the Marathi system while MCD favors the Kannada system, suggesting source-specific artefacts. These are threats to the strength of the attribution, not cases where a 'prediction' is equivalent to an input by construction.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The approach rests on a pre-existing CLS phone inventory, a unified parser, and four FastSpeech2/HiFiGAN synthesisers from prior work. The paper's own hand-chosen components are the phone substitution mappings and the IA/DR parser-rule assignments per target language. No numeric free parameters are fitted in the traditional sense, but the rule assignments are per-language manual choices on which the central claim depends. No new theoretical entities are introduced.

free parameters (2)
  • IA/DR parser-rule assignment per target language = Sanskrit: DR; Maharashtrian Konkani: IA; Canara Konkani: DR; Maithili: IA; Kurukh: IA
    Hand-chosen based on linguistic analysis in Sections 4.3 and 5.2. The claimed improvement depends on these assignments, which are categorical and manually selected, not fitted to audio data.
  • CLS phone substitution mapping for missing sounds = /q/ to /k/, /ɣ/ to /g/, /ɽ/ to /d/, among others listed on external website
    Hand-selected nearest-equivalent mappings in Section 4.2. If these mappings are wrong or incomplete, intelligibility would suffer; the full mapping is not specified in the paper.
assumptions (5)
  • domain assumption The common label set of 72 labels covers all acoustically relevant phones of the source and target languages after substitution mapping.
    Sections 4.1 and 4.2 assume missing sounds can be represented by nearest equivalents without losing contrastive distinctions; no comprehensive inventory analysis is provided.
  • domain assumption The unified parser's IA and DR grapheme-to-phoneme rules are correct and complete for the source languages (Hindi, Marathi, Kannada, Telugu).
    Section 4.1 uses the parser from [36] as-is; any parser errors would propagate into all zero-shot outputs.
  • domain assumption The linguistic characterizations used to assign rules are accurate: Sanskrit lacks schwa deletion, Kurukh has schwa deletion and borrowed IA phonemes, Maharashtrian Konkani is closer to Marathi, Canara Konkani is closer to Kannada.
    Sections 3 and 4.3 cite linguistic sources, but the entire experiment rests on these being correct and transferable to TTS behavior.
  • domain assumption FastSpeech2 and HiFiGAN synthesisers trained on 10 hours of Hindi, Kannada, Marathi, and Telugu are of sufficient quality and suitable for cross-lingual zero-shot use.
    Section 5.1 reports no evaluation of the source synthesisers on their own languages, so cross-language differences could partly reflect source system quality rather than linguistic kinship.
  • domain assumption MOS and WER ratings from a small number of native speakers are reliable indicators of naturalness and intelligibility despite unequal recording conditions and no significance testing.
    Section 5.3 and 5.4 rely on subjective scores from as few as five or six evaluators without confidence intervals or inter-rater reliability measures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Kinship in Speech: Leveraging Linguistic Relatedness for Zero-Shot TTS in Indian Languages." pith.science (2026). https://pith.science/paper/KBSEHJNQ

@misc{pith2026250603884,
  author       = {Pith},
  title        = {Pith review of: Kinship in Speech: Leveraging Linguistic Relatedness for Zero-Shot TTS in Indian Languages},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KBSEHJNQ}},
  note         = {Machine review of arXiv:2506.03884}
}
read the original abstract

Text-to-speech (TTS) systems typically require high-quality studio data and accurate transcriptions for training. India has 1369 languages, with 22 official using 13 scripts. Training a TTS system for all these languages, most of which have no digital resources, seems a Herculean task. Our work focuses on zero-shot synthesis, particularly for languages whose scripts and phonotactics come from different families. The novelty of our work is in the augmentation of a shared phone representation and modifying the text parsing rules to match the phonotactics of the target language, thus reducing the synthesiser overhead and enabling rapid adaptation. Intelligible and natural speech was generated for Sanskrit, Maharashtrian and Canara Konkani, Maithili and Kurukh by leveraging linguistic connections across languages with suitable synthesisers. Evaluations confirm the effectiveness of this approach, highlighting its potential to expand speech technology access for under-represented languages.

Figures

Figures reproduced from arXiv: 2506.03884 by the authors.

Figure 1
Figure 1. Flowchart of the proposed approach: appropriate IA/DR rules are selected based on phonotactics of the target language 4.3. Modifications to text parsing A flowchart of the proposed approach is shown in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Mel spectrogram of Sanskrit word ‘िवकास’ (w-i-k-aa-s-a) in different audio samples: a)Sanskrit GT, b)Hindi synthesised, c)Kannada synthesised, and d)Telugu synthesised. The schwa is preserved (highlighted) in the Kannada and Telugu audio, which is similar to the Sanskrit GT, while it is deleted in the Hindi audio. consider any external linguistic information. As expected, the MOS of Sanskrit synthesis using the Hind… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 51 canonical work pages

  1. [1]

    sprachbund

    Introduction The Indian subcontinent is home to diverse cultures and languages, reflecting its rich linguistic heritage. Unlike the European Union, which primarily uses 3 to 4 major scripts, India has around 66 scripts, of which 13 are used in the 22 official languages. Furthermore, 121 Indian lan- guages are spoken by at least 10,000 individuals [1]. Ind...

    work page

  2. [2]

    sounds in the unseen languages

    We expand the phone-based common label set (CLS) representation for Indian languages [24] to cover missing 1Phonotactics is the sequence of phones allowed in a language. sounds in the unseen languages

    work page

  3. [3]

    We analyse the linguistic and phonotactic properties of the unseen languages and apply the appropriate grapheme-to-phoneme rules

    work page

  4. [4]

    A suitable monolingual TTS system is selected for zero- shot synthesis. Zero-shot synthesis is performed on four languages, collectively spoken by more than 18 million individuals– Sanskrit (classical language), two dialects of Konkani, Maithili and Kurukh (endangered language). The first three are official languages of India. Evaluations demon- strate th...

    work page

  5. [5]

    For example, Kui/Kuvi is written in the Odia script (IA) but belongs to the South-Central Dravidian group

    Motivation Some languages adopt scripts from different language fam- ilies while preserving distinct phonotactic characteristics. For example, Kui/Kuvi is written in the Odia script (IA) but belongs to the South-Central Dravidian group. Bodo is a Sino-Tibetian language and is written in the De- vanagari script (used by IA languages such as Hindi and Marat...

    work page

  6. [6]

    Written in the Devanagari script and despite belonging to the IA language family, Sanskrit does not haveschwa deletion2

    Properties of Sanskrit, Konkani, Maithili and Kurukh Sanskrit, a classical language, has significantly influenced Indian languages. Written in the Devanagari script and despite belonging to the IA language family, Sanskrit does not haveschwa deletion2. Telugu, originating from Proto- Dravidian, has extensive Sanskrit influence, with similar vocabulary and...

    work page

  7. [7]

    Proposed methodology We now discuss the existing framework that supports zero- shot synthesis and present an overview of the proposed methodology. 4.1. Related framework for zero-shot synthesis The number of phonemes in Indian languages is around

    work page

  8. [8]

    Conditional Variational Autoencoder with Ad- versarial Learning for End-to-End Text-to-Speech

    J. Kimet al., “Conditional Variational Autoencoder with Ad- versarial Learning for End-to-End Text-to-Speech.” PMLR, 2021, pp. 5530–5540

    work page 2021

Show all 54 references
  1. [9]

    Next, based on the phonotactics of the target/unseen language, the corresponding family-specific rules are applied

    Instead of directly applying the G2P rules based on the script, the text is first converted to a common represen- tation (grapheme based) using CLS labels. Next, based on the phonotactics of the target/unseen language, the corresponding family-specific rules are applied. For e...

  2. [10]

    Datasets and synthesisers Datasets for training monolingual TTS systems were sourced from [37]6

    Evaluation of zero-shot speech synthesis 5.1. Datasets and synthesisers Datasets for training monolingual TTS systems were sourced from [37]6. Individual FastSpeech2 and HiFiGAN vocoders [3,38,39] were trained for Hindi (Hi), Kannada (Ka), Marathi (Ma) and Telugu (Te) using 10...

  3. [11]

    The difference in dis- tortion between Kannada and Hindi systems for Kurukh is marginal

    is lower compared to the Hindi system and the scores of Telugu and Kannada are similar. The difference in dis- tortion between Kannada and Hindi systems for Kurukh is marginal. Canara Konkani was not considered, as the GT audio was noisy. Notably, the MCD score for Maha- rasht...

  4. [12]

    Conclusions In this work, we were able to generate intelligible and good-quality speech for unseen languages without any adaptation data. Linguistic overlap augmented with a modified shared CLS phone set and parsing rules has made zero-shot synthesis feasible even for language...

  5. [13]

    Office of the registrar general, Paper 1 of 2018 Language, India, States And Union Territories, Table C-16

    I. Office of the registrar general, Paper 1 of 2018 Language, India, States And Union Territories, Table C-16. Delhi: Census of India 2011, 2018. [Online]. Avail- able: https://language.census.gov.in/eLanguageDivision_ VirtualPath/eArchive/pdf/48.pdf

    2018

  6. [14]

    Wavenet: A generative model for raw audio,

    A. van den Oord et al., “Wavenet: A generative model for raw audio,” inSpeech Synthesis Workshop, 2016, p. 125

    2016

  7. [15]

    FastSpeech 2: Fast and High-Quality End-to- End Text to Speech,

    Y. Renet al., “FastSpeech 2: Fast and High-Quality End-to- End Text to Speech,” inICLR, 2021

    2021

  8. [16]

    F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching,

    Y. Chen et al. , “F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching,” 2024. [Online]. Available: https://arxiv.org/abs/2410.06885

    2024 arXiv

  9. [17]

    StyleTTS 2: Towards Human-Level Text- to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models,

    Y. A. Li et al. , “StyleTTS 2: Towards Human-Level Text- to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models,”NeurIPS, vol. 36, pp. 19594–19621, 2023

    2023

  10. [18]

    Glow-TTS: A Generative Flow for Text-to- Speech via Monotonic Alignment Search,

    J. Kim et al. , “Glow-TTS: A Generative Flow for Text-to- Speech via Monotonic Alignment Search,”NeurIPS, vol. 33, pp. 8067–8077, 2020

    2020

  11. [19]

    Grad-TTS: A Diffusion Probabilistic Model for Text-to-Speech,

    V. Popovet al., “Grad-TTS: A Diffusion Probabilistic Model for Text-to-Speech,” vol. 139. PMLR, 2021, pp. 8599–8608

    2021

  12. [20]

    On Achieving and Evaluating Language- Independence in NLP,

    E. M. Bender, “On Achieving and Evaluating Language- Independence in NLP,”Linguistic Issues in Language Tech- nology, vol. 6, pp. 1–28, Oct. 2011

    2011

  13. [21]

    Extending Multilingual Speech Synthesis to 100+ Languages without Transcribed Data,

    T. Saeki et al. , “Extending Multilingual Speech Synthesis to 100+ Languages without Transcribed Data,” inICASSP, 2024, pp. 11546–11550

    2024

  14. [22]

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

    F. Lux et al. , “Meta Learning Text-to-Speech Synthesis in over 7000 Languages,” inInterspeech, 2024, pp. 4958–4962

    2024

  15. [23]

    YourTTS: Towards Zero-Shot Multi- Speaker TTS and Zero-Shot Voice Conversion for everyone

    E. Casanova et al. , “YourTTS: Towards Zero-Shot Multi- Speaker TTS and Zero-Shot Voice Conversion for everyone.” PMLR, 2022, pp. 2709–2720

    2022

  16. [24]

    Speak Foreign Languages with Your Own Voice: Cross-Lingual Neural Codec Language Modeling,

    Z. Zhang et al., “Speak Foreign Languages with Your Own Voice: Cross-Lingual Neural Codec Language Modeling,” arXiv preprint arXiv:2303.03926 , 2023

    2023 arXiv

  17. [25]

    Rapid Speaker Adaptation in Low Resource Text to Speech Systems using Synthetic Data and Transfer learning,

    R. Joshi et al., “Rapid Speaker Adaptation in Low Resource Text to Speech Systems using Synthetic Data and Transfer learning,” in PACLIC 37. ACL, 2023, pp. 267–273

    2023

  18. [26]

    Hierarchical Transfer Learning for Multilin- gual, Multi-Speaker, and Style Transfer DNN-Based TTS on Low-Resource Languages,

    K. Azizahet al., “Hierarchical Transfer Learning for Multilin- gual, Multi-Speaker, and Style Transfer DNN-Based TTS on Low-Resource Languages,”IEEE Access, vol. 8, pp. 179798– 179812, 2020

    2020

  19. [27]

    Uniform Multilingual Multi-Speaker Acoustic Model for Statistical Parametric Speech Synthesis of Low- Resourced Languages

    A. Gutkin, “Uniform Multilingual Multi-Speaker Acoustic Model for Statistical Parametric Speech Synthesis of Low- Resourced Languages.” inInterspeech, 2017, pp. 2183–2187

    2017

  20. [28]

    Survey on the Use of Typological Informa- tion in Natural Language Processing

    O’Horan et al., “Survey on the Use of Typological Informa- tion in Natural Language Processing.” The COLING 2016 Organizing Committee, pp. 1297–1308

    2016

  21. [29]

    Exploring the Role of Language Fam- ilies for Building Indic Speech Synthesisers,

    A. Prakash et al. , “Exploring the Role of Language Fam- ilies for Building Indic Speech Synthesisers,” IEEE/ACM TASLP, vol. 31, p. 734 – 747, 2023

    2023

  22. [30]

    Building Multilingual End-to-End Speech Synthesis- ers for Indian Languages,

    ——, “Building Multilingual End-to-End Speech Synthesis- ers for Indian Languages,” inSSW, 2019, pp. 194–199

    2019

  23. [31]

    Generic Indic Text-to-Speech Synthesisers with Rapid Adaptation in an End-to-End Framework,

    A. Prakash and H. Murthy, “Generic Indic Text-to-Speech Synthesisers with Rapid Adaptation in an End-to-End Framework,” inSSW, 2020, pp. 2962–2966

    2020

  24. [32]

    Exploring the degree of similarities between Hindi and Maithili words from glottochronological perspective,

    S. K. Jha, “Exploring the degree of similarities between Hindi and Maithili words from glottochronological perspective,” International Journal of Innovations in TESOL and Applied Linguistics, vol. 5, 2019

    2019

  25. [33]

    Building African Voices,

    P. Ogayo, G. Neubig, and A. W. Black, “Building African Voices,” in Interspeech, 2022, pp. 1263–1267

    2022

  26. [34]

    India as a lingustic area,

    M. B. Emeneau, “India as a lingustic area,” Language, vol. 32, no. 1, pp. 3–16, 1956. [Online]. Available: http://www.jstor.org/stable/410649

    1956

  27. [35]

    Proposition 16,

    N. Trubetzkoy, “Proposition 16,” inActes du Premier Con- grès International de Linguistes, a La Haye . W. Sijthoff, 1930, 10–15 Apr 1928, pp. 17–18

    1930

  28. [36]

    A common attribute based unified HTS framework for speech synthesis in Indian languages,

    B. Ramani et al., “A common attribute based unified HTS framework for speech synthesis in Indian languages,” inSSW, 2013, pp. 291–296

    2013

  29. [37]

    Masica, The Indo-Aryan Languages , ser

    C. Masica, The Indo-Aryan Languages , ser. Cambridge Language Surveys. Cambridge University Press, 1993. [Online]. Available: https://books.google.co.in/books?id= J3RSHWePhXwC

    1993

  30. [38]

    Krishnamurti, The Dravidian Languages , ser

    B. Krishnamurti, The Dravidian Languages , ser. Cambridge Language Surveys. Cambridge University Press, 2003. [Online]. Available: https://books.google.co.in/books?id= 54fV7Lwu3fMC

    2003

  31. [39]

    Technology Pipeline for Large Scale Cross-Lingual Dubbing of Lecture Videos into Multiple In- dian Languages,

    A. Prakash et al. , “Technology Pipeline for Large Scale Cross-Lingual Dubbing of Lecture Videos into Multiple In- dian Languages,” inInterspeech, 2023, pp. 3683–3684

    2023

  32. [40]

    BhashaVerse : Translation Ecosystem for Indian Subcontinent Languages,

    V. Mujadia and D. M. Sharma, “BhashaVerse : Translation Ecosystem for Indian Subcontinent Languages,” 2025. [Online]. Available: https://arxiv.org/abs/2412.04351

    2025 arXiv

  33. [41]

    Comparative Analysis of Tel- ugu and Sanskrit Languages,

    T. K. Rao and T. V. Prased, “Comparative Analysis of Tel- ugu and Sanskrit Languages,”KAA V International Journal of Science, Engineering Technology , 2014

    2014

  34. [42]

    Kannada versus Sanskrit: Hegemony, Power and Subjugation,

    M. Mallikarjun, “Kannada versus Sanskrit: Hegemony, Power and Subjugation,” Language in India , vol. 17, no. 8, 2017

    2017

  35. [43]

    Empirical Analysis of Oral and Nasal Vowels of Konkani,

    S. Fadte et al. , “Empirical Analysis of Oral and Nasal Vowels of Konkani,” 2023. [Online]. Available: https://arxiv.org/abs/2305.10122

    2023 arXiv

  36. [45]

    Kobayashi and B

    M. Kobayashi and B. Tirkey,The Kurux Language: Gram- mar, Texts, and Lexicon , ser. Brill’s Studies in South and Southwest Asian Languages, 2017, vol. 08, no. 8

    2017

  37. [46]

    Languages in Contact in Jharkhand. A Case of Language Conflation, Language Change and Language Con- vergence,

    A. Abbi, “Languages in Contact in Jharkhand. A Case of Language Conflation, Language Change and Language Con- vergence,” in Languages of tribal and indigenous peoples of India: the ethnic space , 1997, no. 10, pp. 131––148

    1997

  38. [47]

    Code-switching in Indic Speech Syn- thesisers,

    A. L. Thomas et al., “Code-switching in Indic Speech Syn- thesisers,” in Interspeech, 2018, pp. 1948–1952

    2018

  39. [48]

    A Unified Parser for Developing Indian Lan- guage Text to Speech Synthesizers,

    A. Babyet al., “A Unified Parser for Developing Indian Lan- guage Text to Speech Synthesizers,” inTSD, 2016, pp. 514– 521

    2016

  40. [49]

    Resources for Indian languages,

    ——, “Resources for Indian languages,” in TSD, 2016, pp. 37–43

    2016

  41. [50]

    The common label set (CLS) [24] provides a set of labels, wherein acoustically similar phones across 13 Indian lan- guages are mapped together

    There are 15-18 vowels and 35-38 consonants [35]. The common label set (CLS) [24] provides a set of labels, wherein acoustically similar phones across 13 Indian lan- guages are mapped together. The CLS, with a token size of 72, provides a very compact representation for train-...

  42. [51]

    Towards Developing State-of-The-Art TTS Synthesisers for 13 Indian Languages with Signal Pro- cessing Aided Alignments,

    A. Prakash et al. , “Towards Developing State-of-The-Art TTS Synthesisers for 13 Indian Languages with Signal Pro- cessing Aided Alignments,” inASRU, 2023, pp. 1–8

    2023

  43. [52]

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

    J. Kong et al., “HiFi-GAN: generative adversarial networks for efficient and high fidelity speech synthesis,” inNeurIPS, 2020, pp. 17022–17033

    2020

  44. [53]

    VoiceFixer: A Unified Framework for High- Fidelity Speech Restoration,

    H. Liu et al. , “VoiceFixer: A Unified Framework for High- Fidelity Speech Restoration,” inInterspeech, 2022, pp. 4232– 4236

    2022

  45. [54]

    The SUS test: A method for the assessment of text-to-speech synthesis intelligibility using Semantically Unpredictable Sentences,

    C. Benoîtet al., “The SUS test: A method for the assessment of text-to-speech synthesis intelligibility using Semantically Unpredictable Sentences,” Speech Communication, vol. 18, no. 4, pp. 381–392, 1996

    1996

  46. [55]

    Mel-cepstral distance measure for objective speech quality assessment,

    R. Kubichek, “Mel-cepstral distance measure for objective speech quality assessment,” inIEEE Pacific Rim Conference on Communications Computers and Signal Processing, vol. 1, 1993, pp. 125–128

    1993

Pith tools

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