ParsVoice: A Large-Scale Multi-Speaker Persian Speech Corpus for Text-to-Speech Synthesis

Azadeh Shakery; Heshaam Faili; Mohammad Javad Ranjbar Kalahroodi

REVIEW 4 major objections 3 minor 28 references

ParsVoice: A Large-Scale Multi-Speaker Persian Speech Corpus for Text-to-Speech Synthesis

T0 review · 4 major / 3 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read ParsVoice, built automatically from 2,000 audiobooks, is claimed to be the largest openly available Persian speech-text corpus for multi-speaker TTS, with thousands of aligned hours.

desk verdict A genuinely useful Persian TTS corpus, but the abstract's headline numbers contradict the paper's own tables, so the scale claim needs a correction before it can be trusted. read the letter →

arxiv 2510.10774 v3 pith:UERMLUC4 submitted 2025-10-12 cs.SD cs.AIcs.HCcs.LG

Mohammad Javad Ranjbar Kalahroodi , Heshaam Faili , Azadeh Shakery This is my paper

classification cs.SDcs.AIcs.HCcs.LG

keywords PersianTTScorpusmulti-speakertext-to-speechaudiobookprocessingspeech-textalignmentsentencecompletiondetectionspeakerdiarizationlow-resourcelanguagespeechconstruction

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

The paper introduces ParsVoice, a large-scale Persian speech-text corpus assembled from public audiobooks via an automated pipeline, and argues it is the largest openly available resource of its kind for multi-speaker text-to-speech. The pipeline turns long-form audio into short, sentence-aligned segments using a sentence-completeness classifier, ASR-driven boundary trimming, and Persian-specific quality filters. The release described ranges from roughly 1,800 hours in the main analysis to 2,200 TTS-ready hours in the abstract, with more than 1,800 automatically identified speakers. The authors validate the corpus by fine-tuning a zero-shot multilingual TTS model, obtaining naturalness and speaker-similarity MOS scores of 3.6 and 4.0 out of 5. If the claims hold, Persian TTS research gains a data resource comparable in scale to major English corpora.

What carries the argument

The central mechanism is the automated corpus-construction pipeline. Its load-bearing piece is the interplay between automatic speech recognition transcription and a fine-tuned Persian sentence-completion classifier: candidate audio segments are transcribed, the classifier checks whether the text is a complete sentence, and segments that fail receive stepwise boundary extensions up to 5 seconds with retranscription until they pass or are discarded. Around this core sit a binary-search boundary optimizer that trims leading and trailing silence by verifying that retranscription remains stable, composite quality scores for both text (character quality, length, repetition, phonetic coverage) and

What would settle it

Select a random sample of, say, 200 segments from the released corpus, have a fluent Persian speaker transcribe the audio, and compare to the released text; if the segment-level word error rate is high (e.g., above 20%) or many released texts are not complete sentences, the corpus's TTS-readiness claim would be undermined.

Watch

Extended reading notes

Core claim

ParsVoice is a new Persian speech corpus released for public use, generated by a fully automated pipeline that converts long-form audiobook audio into short, sentence-aligned segments. The paper's central assertion is that this corpus, at roughly 1,800 hours (the abstract cites 2,200 TTS-ready hours) and over 1,800 automatically identified speakers, is the largest openly available Persian TTS corpus to date, more than 25 times larger than the previous leader. Supporting the corpus is a pipeline whose key components are a sentence-completeness classifier, an ASR-based binary-search boundary trimmer, Persian-specific text and audio quality scorers, and a two-stage speaker identification step.

Load-bearing premise

The entire corpus's usability as ground-truth text-to-speech data rests on the automatic speech recognition transcriptions being accurate enough to treat as the text, since the original audiobook text was not used and no human transcription-accuracy audit of the corpus is reported.

Editorial extensions

If this is right

Persian multi-speaker TTS training no longer needs to rely on single-speaker corpora under 100 hours; a corpus of roughly 1,800 hours with hundreds of speakers is now available.
The pipeline demonstrates a scalable route from publicly available audiobooks to TTS-ready data, which could be adapted for other low-resource languages.
A zero-shot TTS model fine-tuned on ParsVoice achieves competitive MOS scores, showing the corpus can support natural Persian voice synthesis without phoneme representations.
The dataset's scale and speaker diversity enable research in speaker adaptation and voice cloning for Persian, areas previously limited by tiny single-speaker datasets.
The clean, aligned segments may also serve as training data for other speech tasks such as automatic speech recognition.

Reading between the lines

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

Because transcriptions come from a cloud ASR service rather than the original book text or human review, the corpus's ceiling is set by ASR accuracy; a small human audit of random segments would establish its true word error rate and should be the first thing a downstream user checks.
The reported speaker count is ambiguous: over 1,815 automatically identified speaker instances in the full data, but 470+ speakers in the filtered subset; the 97% agreement with known narrator labels suggests clustering quality but not perfect identity merging.
The same pipeline likely transfers to other under-resourced languages with audiobook sources, provided a sentence-completeness model and a decent ASR exist; this could accelerate low-resource TTS data creation globally.
The discrepancy between the abstract's 2,200 hours and the body's 1,804 hours suggests the release may contain both a broader and a stricter filtered subset; users should verify which one they download and whether it matches their quality threshold.

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.

Desk Editor's Note

read the letter

ParsVoice is a real contribution: the first open Persian TTS corpus at this scale, 1,804 hours from 470+ speakers, a step change over ManaTTS (86h) and DeepMine-Multi-TTS (120h). The pipeline is sensible, the speaker ID work is reasonable, and the XTTS fine-tune shows the data can train a working multi-speaker model. That part is solid and worth building on.

The problem is the numbers. The abstract says 2,200 hours, 1.36M segments, 1,815 speakers, and “more than 25 times larger” than prior open Persian TTS data. The body — Table 1, Table 2, and the conclusion — says 1,804 hours, 1.15M segments, and 470+ speakers. That is a ~22% hour gap, an ~18% segment gap, and roughly a 4x speaker-count gap. The “25x” claim doesn't hold with either set of numbers against the cited 120h DeepMine-Multi-TTS (closer to 15x or 18x). This is not a cosmetic discrepancy; the whole headline claim is quantitative. It must be reconciled before the corpus's scale can be evaluated.

The deeper scientific risk is the transcripts. The text labels are Google Speech-to-Text output, not original audiobook text or manually verified captions. The paper reports no transcription-accuracy audit on the corpus itself. For TTS, wrong text is poison; a small human-checked sample of, say, a few hundred segments would address this directly. The MOS numbers would also benefit from confidence intervals — 40 raters and 90 samples is fine, but report the variance. Minor: pipeline code is not released, only the dataset, so replication is harder.

Overall, this is a useful resource with a fixable presentation flaw and a real but addressable ASR-transcript concern. The authors appear to know what they are doing; the internal inconsistency reads as a rushed write-up rather than deliberate misdirection. I would send it to peer review — a good referee can push for the number reconciliation and a text-quality audit. If you work on low-resource TTS, read it; if not, it's a maybe for the reading group.

Referee Report

4 major / 3 minor

Summary. The manuscript presents ParsVoice, an automatically constructed Persian speech-text corpus for TTS, derived from 2,000 audiobooks from IranSeda. The construction pipeline consists of VAD segmentation, Google ASR transcription, ParsBERT-based sentence completeness filtering, binary-search boundary trimming, Persian text/audio quality scoring, ECAPA-TDNN speaker diarization/global clustering, and punctuation restoration. The authors claim that this is the largest publicly available Persian TTS corpus and validate it by fine-tuning XTTS, reporting MOS=3.6, SMOS=4.0. The body reports 3,526.4 hours raw, 1,803.9 hours and 1,147,718 segments after TTS filtering, and 470+ speakers; the stand-alone abstract reports 2,200 hours, 1.36M segments, and 1,815 speaker IDs.

Significance. If the headline statistics are corrected and transcript accuracy is verified, ParsVoice would be a substantial resource: it offers roughly an order-of-magnitude scale-up over the largest open Persian TTS corpus, has broad speaker diversity (470+), and is publicly released. The pipeline description is detailed enough to replicate; the authors also compare with previous Persian TTS datasets and provide both subjective and objective evaluation. However, the current internal inconsistencies and the absence of a text-quality audit make the central 'largest / TTS-ready' claim impossible to verify from the manuscript.

major comments (4)

[Abstract vs. §4.1, Table 1, Table 2, §6] The headline statistics in the stand-alone abstract — 2,200 h, 1.36M segments, 1,815 speaker IDs, and '>25x larger' — are not supported by the body. Table 1 shows 1,803.9 h and 1,147,718 segments after TTS filtering; Table 2 and §6 report 1,804 h and 470+ speakers. §4.1 says 'over 1,815 unique speaker instances across the entire dataset,' not in the TTS-ready subset. The '25x' claim is also undefined: against DeepMine-Multi-TTS (120 h), 2,200 h is ~18x and 1,804 h is ~15x; against ManaTTS (86 h), 1,804 h is ~21x, but ManaTTS is not the previous largest TTS corpus. Because the paper's central claim is quantitative, this contradiction must be resolved: either add table rows for the 2,200 h subset or correct the abstract.
[§3.2 Phase 2; §5] The corpus text is entirely produced by Google Speech-to-Text ASR (§3.2, Phase 2), then punctuated by a model. No manual or reference-based transcription accuracy measure is reported for the corpus itself. The WER/CER reported in §5 (22.57%/12.78%) is computed on synthesized speech, not on the corpus text labels. Since TTS training uses these labels as ground truth, the paper should report a word/character error rate or a manual audit on a random sample of corpus segments against either the original audiobook text or human transcriptions, broken down by speaker and quality score. Without this, the 'TTS-ready' claim is not supported.
[§5, Table 3] Subjective scores are reported as point estimates (MOS 3.60, SMOS 4.00) with no confidence intervals or statistical comparison to baselines. With 90 synthesized sentences and 40 raters, this is not a small evaluation, but CIs are essential before claiming 'competitive performance.' In addition, the FastSpeech2 baselines in Table 3 come from a different paper ([12]) with a different evaluation protocol, so the comparison is only indicative; either re-run the baselines under the same protocol or soften the claim.
[§3.4, §3.6] The quality filtering thresholds (audio score ≥0.8, text score ≥0.5) and the 'empirical weights' for text and audio metrics are stated without values or justification. Since these thresholds define the released 1,804-hour subset and the TTS-ready claim, the authors should publish the exact scoring formulas/weights and ideally report sensitivity of the final size/quality to the thresholds. This is not a stylistic point: the reproducibility of the corpus statistics depends on it.

minor comments (3)

[Throughout] There are several typos and formatting artifacts: 'ParsV oice', 'V oice', 'identify and filter sentence fragments' lacks a subject, and the dataset URL appears as 'ParsVoice' without a hyperlink. These should be cleaned before resubmission.
[§4.1] The linguistic statistics (267,965 unique words from 25,499,474 tokens) are given without defining tokenization or whether punctuation was removed. Please specify the tokenization method.
[§6] The conclusion mentions 'an additional 2,000 hours of high-quality speech' beyond the TTS subset, but Table 1 has only 'before' and 'after TTS filtering' rows. Clarify whether these 2,000 hours come from the 3,526.4-hour 'before' set and how they relate to the 1,803.9-hour subset.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ParsVoice's corpus construction and evaluation are self-contained; the only author-overlap citation (PersianPunc) supports auxiliary pipeline components and is not load-bearing.

full rationale

The paper's central output is a corpus, not a derived quantity. Construction uses external tools (ParsBERT, WebRTC, Google STT, ECAPA-TDNN, inaSpeechSegmenter) and hand-set quality thresholds; validation is an XTTS fine-tune evaluated by human MOS and objective similarity. No reported statistic reduces by construction to a fitted input, and no evaluation number is the value of a pipeline parameter. The only author-overlapping citation, [14] PersianPunc, is used to train sentence-completion and punctuation-restoration components; the central 'largest Persian TTS corpus' claim does not rest on it. The abstract's 2,200-hour/1.36M-segment/1,815-speaker numbers conflict with the paper's own Table 1 (1,803.9 h; 1,147,718 segments) and Table 2/conclusion (1,804 h; 470+ speakers), and the '25x' comparison is not reproducible from those tables; this is a quantitative consistency/verifiability problem, not circularity. Similarly, using Google STT transcripts as text labels without a corpus-level transcription audit is a data-quality risk. No circular step is exhibited.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on hand-set quality thresholds, empirical weights, and strong assumptions about ASR and speaker-clustering accuracy. None of these are reported numerically or audited against a manual gold standard, so they are the main unmeasured inputs to the corpus.

free parameters (6)

empirical weights for Persian text quality metrics = not reported
§3.4.1: metrics are 'aggregated using empirical weights' to compute the total score; no values or sensitivity analysis are given.
empirical weights for audio quality metrics = not reported
§3.4.2: the composite audio quality score uses 'empirically weighted' standard audio metrics; weights are never specified.
TTS filtering thresholds = audio ≥ 0.8, text ≥ 0.5
§3.6: final cleaning thresholds are chosen without ablation or reported justification.
initial boundary trim duration = 3 seconds
§3.3: 'remove 3 seconds from both the beginning and end' before binary search; no justification for this value.
boundary search increments = 0.1 seconds, up to 5 seconds
§3.2 Phase 3 and §3.3: iterative boundary extensions and linear fine-tuning use 0.1-second steps without sensitivity analysis.
XTTS fine-tuning hyperparameters = batch size 16, 170,000 steps, 2,500 new tokens
§5: training choices are reported but no ablations or tuning curves are provided.

assumptions (4)

domain assumption Google Speech-to-Text produces sufficiently accurate Persian transcriptions to serve as ground-truth TTS text.
§3.2 Phase 2: every text label comes from this ASR; no manual audit of corpus transcription accuracy is reported.
domain assumption IranSeda audiobooks are free of copyright restrictions and can be redistributed.
§3.1: asserted without license documentation or legal verification.
domain assumption ECAPA-TDNN embeddings and the described clustering pipeline recover true speaker identities.
§3.5: validated against metadata on a subset (97% consistency), but 40% of books lack narrator metadata, so the global speaker count is an estimate.
domain assumption ASR-based boundary optimization preserves exact sentence boundaries and alignment.
§3.3: alignment quality is inferred from transcription stability, not verified against human annotations.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ParsVoice: A Large-Scale Multi-Speaker Persian Speech Corpus for Text-to-Speech Synthesis." pith.science (2026). https://pith.science/paper/UERMLUC4

@misc{pith2026251010774,
  author       = {Pith},
  title        = {Pith review of: ParsVoice: A Large-Scale Multi-Speaker Persian Speech Corpus for Text-to-Speech Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UERMLUC4}},
  note         = {Machine review of arXiv:2510.10774}
}

read the original abstract

Persian remains substantially underrepresented in open speech-text resources, limiting progress in multi-speaker text-to-speech (TTS), speech-language modelling, and low-resource speech processing. We introduce ParsVoice, the largest publicly available Persian speech-text corpus tailored for training multi-speaker TTS systems, along with a scalable pipeline to construct high-quality speech-text data from long-form audiobook recordings. The pipeline combines a fine-tuned ParsBERT sentence-completion classifier, ASR-based boundary optimization, punctuation restoration, speaker identification, and a multi-dimensional quality assessment that covers both audio and Persian-specific text properties. The resulting release contains a 2,200-hour TTS-ready subset with 1.36 million aligned segments from 1,815 automatically identified speaker IDs, making it more than 25 times larger than the previously largest open Persian TTS dataset. To validate the corpus, we fine-tune XTTS, a zero-shot multilingual TTS model that operates directly on raw Persian text without phoneme representations, achieving a naturalness MOS of 3.6/5 and speaker similarity MOS of 4.0/5. The ParsVoice dataset is publicly available at: https://huggingface.co/datasets/MohammadJRanjbar/ParsVoice.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 7 linked inside Pith

[12]

Librispeech: An asr corpus based on public domain audio books

Panayotov, Vassil, Chen, Guoguo, Povey, Daniel, and Khudanpur, Sanjeev. Librispeech: An asr corpus based on public domain audio books. In2015 IEEE Interna- tional Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), pages 5206–5210, 2015

2015
[1]

Persian, although spoken by more than 100 million people worldwide, remains significantly under- represented in speech corpora compared to high-resource languages such as English

INTRODUCTION As transformer architectures [1] and generative models, rapidly advance, the scarcity of high-quality training data has become even more pronounced, especially for low- resource languages. Persian, although spoken by more than 100 million people worldwide, remains significantly under- represented in speech corpora compared to high-resource la...
[2]

While multilingual efforts like Common V oice arXiv:2510.10774v2 [cs.SD] 14 Oct 2025 Fig

RELATED WORK Speech dataset development has been dominated by En- glish resources such as LibriSpeech [2], LJSpeech [3], and VCTK [4]. While multilingual efforts like Common V oice arXiv:2510.10774v2 [cs.SD] 14 Oct 2025 Fig. 1. Overview of the proposed pipeline. [5], Multilingual LibriSpeech [6], and V oxPopuli [7] have expanded to 20+ languages, they rem...

arXiv 2025
[3]

AUTOMATED CORPUS CONSTRUCTION PIPELINE We introduce a pipeline that transforms raw audiobook recordings into a structured, high-quality speech and text corpus through interconnected stages designed to maximize data quality. The pipeline addresses key challenges in Persian speech data creation: maintaining sentence integrity, ensuring audio-text alignment ...
[4]

Initial Adjustment:We start by removing 3 seconds from both the beginning and end of each segment, then per- form re-transcription
[5]

Stability Verification:If the new transcription differs significantly from the original, the trimming is deemed exces- sive and must be reduced
[6]

Binary search continues until re-transcription differs from the origi- nal

Binary Search Optimization:The algorithm employs binary search, iteratively halving the trimming interval to ef- ficiently converge on the optimal boundary position. Binary search continues until re-transcription differs from the origi- nal
[7]

This approach ensures that each segment contains only the essential speech content while maintaining transcription accuracy

Fine-Grained Linear Search:After binary search converges, a linear search with 0.1-second increments is ap- plied to achieve precise boundary alignment where further bi- nary search refinement is no longer effective. This approach ensures that each segment contains only the essential speech content while maintaining transcription accuracy. 3.4. Text-Audio...

Show all 28 references

[8]

Collection and Processing Results Out of 3,807 books (9,538 hours), we fully processed 2,000

PARSVOICE CORPUS ANALYSIS 4.1. Collection and Processing Results Out of 3,807 books (9,538 hours), we fully processed 2,000. Our automated pipeline generated 5,158,344 initial audio seg- ments. After removing empty segments, 3,321,212 segments remained. Table 2. ParsV oice Dat...
[9]

Training:A BPE model was trained, and 2,500 new Per- sian tokens were extracted from Copera and added to the GPT model vocabulary

EV ALUATION: TTS MODEL TRAINING To validate ParsV oice for TTS applications, we fine-tuned XTTS [18], a state-of-the-art multi-lingual TTS model with zero-shot capabilities. Training:A BPE model was trained, and 2,500 new Per- sian tokens were extracted from Copera and added t...
[10]

CONCLUSION In this work, we address the scarcity of high-quality Persian speech datasets by introducing ParsV oice, the largest publicly available Persian dataset to date. ParsV oice consists of 1,804 hours of clean, segmented speech from 470+ distinct speakers suitable for TT...
[11]

Attention is all you need

Vaswani, Ashish, Shazeer, Noam, Parmar, Niki, Uszko- reit, Jakob, Jones, Llion, Gomez, Aidan N., Kaiser, Lukasz, and Polosukhin, Illia. Attention is all you need. Advances in Neural Information Processing Systems, pages 5998–6008, 2017

2017
[13]

The lj speech dataset, 2017.https://keithito.com/ LJ-Speech-Dataset/

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

2017
[14]

Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit, 2017

Veaux, Christophe, Yamagishi, Junichi, and MacDon- ald, Kirsten. Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit, 2017. The Univer- sity of Edinburgh. Centre for Speech Technology Re- search (CSTR)

2017
[15]

Common voice: A massively-multilingual speech corpus, 2020

Ardila, Rosana, Branson, Megan, Davis, Kelly, Hen- retty, Michael, Kohler, Michael, Meyer, Josh, Morais, Reuben, Saunders, Lindsay, Tyers, Francis M., and We- ber, Gregor. Common voice: A massively-multilingual speech corpus, 2020. arXiv:1912.06670

2020 arXiv
[16]

Mls: A large- scale multilingual dataset for speech research

Pratap, Vineel, Xu, Qiantong, Sriram, Anuroop, Syn- naeve, Gabriel, and Collobert, Ronan. Mls: A large- scale multilingual dataset for speech research. InInter- speech 2020, October 2020

2020
[17]

V oxpopuli: A large-scale multilingual speech corpus for representation learn- ing, semi-supervised learning and interpretation, 2021

Wang, Changhan et al. V oxpopuli: A large-scale multilingual speech corpus for representation learn- ing, semi-supervised learning and interpretation, 2021. arXiv:2101.00390

2021 arXiv
[18]

A multi purpose and large scale speech corpus in persian and english for speaker and speech recognition: the deepmine database, 2019.https: //arxiv.org/abs/1912.03627

Zeinali, Hossein, Burget, Luk ´aˇs, and ˇCernock´y, Jan “Honza”. A multi purpose and large scale speech corpus in persian and english for speaker and speech recognition: the deepmine database, 2019.https: //arxiv.org/abs/1912.03627

2019 arXiv
[19]

Ar- mantts single-speaker persian dataset.arXiv preprint arXiv:2304.03585, 2023

Shamgholi, Mohammd Hasan, Saeedi, Vahid, Peyman- fard, Javad, Alhabib, Leila, and Zeinali, Hossein. Ar- mantts single-speaker persian dataset.arXiv preprint arXiv:2304.03585, 2023

2023 arXiv
[20]

Manatts persian: a recipe for creating tts datasets for lower resource languages.arXiv preprint arXiv:2409.07259, 2024

Fetrat Qharabagh, Mahta, Dehghanian, Zahra, and Ra- biee, Hamid R. Manatts persian: a recipe for creating tts datasets for lower resource languages.arXiv preprint arXiv:2409.07259, 2024

2024 arXiv
[21]

End-to-end multi- speaker fastspeech2 with hierarchical decoder.IEEE Ac- cess, 13:127805–127814, 2025

Adibian, Majid and Zeinali, Hossein. End-to-end multi- speaker fastspeech2 with hierarchical decoder.IEEE Ac- cess, 13:127805–127814, 2025

2025
[22]

Deepmine-multi-tts: a persian speech corpus for multi-speaker text-to-speech.Language Resources and Evaluation, 59:2245–2264, 2025

Adibian, Majid, Zeinali, Hossein, and Barmaki, Soroush. Deepmine-multi-tts: a persian speech corpus for multi-speaker text-to-speech.Language Resources and Evaluation, 59:2245–2264, 2025

2025
[23]

Parsbert: Transformer-based model for persian language under- standing.Neural Processing Letters, 53(6):3831–3847, October 2021

Farahani, Mehrdad, Gharachorloo, Mohammad, Fara- hani, Marzieh, and Manthouri, Mohammad. Parsbert: Transformer-based model for persian language under- standing.Neural Processing Letters, 53(6):3831–3847, October 2021

2021
[24]

Persianpunc, 2025

Ranjbar, Mohammad J., Shakery, Azadeh, and Faili, Heshaam. Persianpunc, 2025. Hugging Face.https://huggingface.co/datasets/ MohammadJRanjbar/PersianPunc

2025
[25]

Webrtc voice activity detector.https: //github.com/wiseman/py-webrtcvad

Google. Webrtc voice activity detector.https: //github.com/wiseman/py-webrtcvad. Ac- cessed: 2025-09-14

2025
[26]

Ina’s mirex 2018 music and speech detec- tion system

Doukhan, David, Lechapt, Eliott, Evrard, Marc, and Carrive, Jean. Ina’s mirex 2018 music and speech detec- tion system. InMusic Information Retrieval Evaluation eXchange (MIREX 2018), 2018

2018
[27]

ECAPA-TDNN: Empha- sized channel attention, propagation and aggregation in TDNN based speaker verification

Desplanques, Brecht et al. ECAPA-TDNN: Empha- sized channel attention, propagation and aggregation in TDNN based speaker verification. In Meng, He- len et al., editors,Interspeech 2020, pages 3830–3834. ISCA, 2020

2020
[28]

Xtts: a massively multilingual zero-shot text-to-speech model, 2024

Casanova, Edresson, Davis, Kelly, G¨olge, Eren, G¨oknar, G¨orkem, Gulea, Iulian, Hart, Logan, Aljafari, Aya, Meyer, Joshua, Morais, Reuben, Olayemi, Samuel, and Weber, Julian. Xtts: a massively multilingual zero-shot text-to-speech model, 2024. arXiv:2406.04904

2024 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [12]

Librispeech: An asr corpus based on public domain audio books

Panayotov, Vassil, Chen, Guoguo, Povey, Daniel, and Khudanpur, Sanjeev. Librispeech: An asr corpus based on public domain audio books. In2015 IEEE Interna- tional Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), pages 5206–5210, 2015

2015

[2] [1]

Persian, although spoken by more than 100 million people worldwide, remains significantly under- represented in speech corpora compared to high-resource languages such as English

INTRODUCTION As transformer architectures [1] and generative models, rapidly advance, the scarcity of high-quality training data has become even more pronounced, especially for low- resource languages. Persian, although spoken by more than 100 million people worldwide, remains significantly under- represented in speech corpora compared to high-resource la...

[3] [2]

While multilingual efforts like Common V oice arXiv:2510.10774v2 [cs.SD] 14 Oct 2025 Fig

RELATED WORK Speech dataset development has been dominated by En- glish resources such as LibriSpeech [2], LJSpeech [3], and VCTK [4]. While multilingual efforts like Common V oice arXiv:2510.10774v2 [cs.SD] 14 Oct 2025 Fig. 1. Overview of the proposed pipeline. [5], Multilingual LibriSpeech [6], and V oxPopuli [7] have expanded to 20+ languages, they rem...

arXiv 2025

[4] [3]

AUTOMATED CORPUS CONSTRUCTION PIPELINE We introduce a pipeline that transforms raw audiobook recordings into a structured, high-quality speech and text corpus through interconnected stages designed to maximize data quality. The pipeline addresses key challenges in Persian speech data creation: maintaining sentence integrity, ensuring audio-text alignment ...

[5] [4]

Initial Adjustment:We start by removing 3 seconds from both the beginning and end of each segment, then per- form re-transcription

[6] [5]

Stability Verification:If the new transcription differs significantly from the original, the trimming is deemed exces- sive and must be reduced

[7] [6]

Binary search continues until re-transcription differs from the origi- nal

Binary Search Optimization:The algorithm employs binary search, iteratively halving the trimming interval to ef- ficiently converge on the optimal boundary position. Binary search continues until re-transcription differs from the origi- nal

[8] [7]

This approach ensures that each segment contains only the essential speech content while maintaining transcription accuracy

Fine-Grained Linear Search:After binary search converges, a linear search with 0.1-second increments is ap- plied to achieve precise boundary alignment where further bi- nary search refinement is no longer effective. This approach ensures that each segment contains only the essential speech content while maintaining transcription accuracy. 3.4. Text-Audio...