CVSS-X: A Multilingual Speech-to-Speech Translation Corpus for 28 Languages
Lucas Rafael Stefanel Gris, Alef Iury Siqueira Ferreira, Frederico Santos de Oliveira, Augusto Seben da Rosa, Alexandre Costa Ferro Filho, Arlindo Rodrigues Galvão Filho, Anderson da Silva Soares
cs.CL · cs.SD · eess.AS
Quick verdict
Large EN→28-language synthetic S2ST corpus (16k h, two TTS variants); solid resource work, but its 'comparable quality to CVSS' claim exceeds Tables 4-5.
Summary
CVSS-X reverses the translation direction of CVSS to build a synthetic speech-to-speech translation corpus from English into 28 target languages across seven typological families. It provides about 240,000 parallel speech pairs per language (240,192; over 6.7 million pairs and roughly 16,000 hours total) in two variants: CVSS-X-C with two fixed reference voices per language, and CVSS-X-T with cross-lingual voice cloning. Evaluation on 200 dev utterances per language with Whisper large-v3 (WER/CER, ASR-BLEU) and UTMOS naturalness shows the canonical variant at 12.1% WER and 82.4 ASR-BLEU, with clearly lower naturalness than the English-only CVSS (UTMOS 3.55 versus 4.43); the transferred variant scores lower still, with the largest BLEU drop on Chinese. Acknowledged limitations include no human listening test, a translation-model choice backed by only 100 closed-set English-Portuguese LLM-judged examples, a dev set supplemented from training samples without documented deduplication, and no downstream S2ST model validation yet.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 19 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- CVSS-X is the first large-scale, fully downloadable one-to-many (EN→X) S2ST corpus, covering 28 languages in 12 families (7 macro-categories) with ~240,192 parallel pairs per language and over 16,000 hours.
- Because target speech is synthesized from verified parallel text, the corpus offers perfect sentence-level alignment, unlike mined corpora such as SpeechMatrix or SeamlessAlign.
- Its translation quality and cross-family consistency are comparable to CVSS, enabling bidirectional and pivot-based X→Y S2ST research when combined with CVSS.
- Voice cloning (CVSS-X-T) preserves speaker identity with ECAPA-TDNN similarity of 0.607 at a modest naturalness cost (UTMOS 3.21 vs 3.55 for canonical).
Strengths
First fully downloadable large-scale EN→X speech corpus across 28 languages
Section 2/Table 1 position CVSS-X against mined corpora (SpeechMatrix 418K h, 17 European languages; SeamlessAlign metadata-only): CVSS-X offers 6,725,376 pairs (~16,070 h per Table 3) in 28 languages and 7 families, with perfect sentence-level alignment because target speech is synthesized from verified parallel text.
Deliberate design symmetry with CVSS
Section 3.1 recovers 240,192 of 264,037 Common Voice v4 English utterances (91.0%) used by CVSS/CoVoST 2 via normalized text matching, so train/dev/test sizes (222,349/10,000/7,843) mirror the original corpus and enable bidirectional studies when combined with CVSS.
Multi-dimensional automated evaluation with sensible language-specific handling
Statistically grounded sample sizing
Transparent limitations and licensing discussion
Weaknesses
Abstract's 'comparable translation quality' contradicts its own tables
Translation-model selection is thin for a 28-language corpus
Dev/train contamination risk in the evaluation split
No downstream S2ST model validation
Fully automated quality metrics without anchors
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
CVSS-X Review — A 16,000-hour EN→X Synthetic S2ST Corpus
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the CVSS-X dataset and how does it differ from CVSS?
CVSS-X is a synthetic speech-to-speech translation corpus translating from English into 28 target languages, reversing the direction of the original CVSS corpus (21 languages into English). It contains about 240,192 parallel speech pairs per language — over 6.7 million pairs, roughly 16,000 hours — and comes in two variants: CVSS-X-C with two fixed reference voices per language, and CVSS-X-T that clones the original English speaker's voice into the target language.
Is CVSS-X free to use commercially?
No. The dataset is released under CC-BY-NC 4.0, a non-commercial license the authors inherit from the NLLB-200 translation model used to produce the text. The paper notes that a future regeneration using TranslateGemma-12B could enable a more permissive Apache 2.0 license.
How good is the quality of CVSS-X compared to CVSS?
Using identical evaluation (Whisper large-v3 round-trip ASR-BLEU, WER/CER, UTMOS), CVSS-X-C reaches 82.4 ASR-BLEU and 12.1% WER versus CVSS-C's 94.2 and 3.5%, with naturalness of 3.55 UTMOS versus 4.43. Quality is consistent across language families but clearly below the English-only CVSS; the cloned-voice variant scores lower still.
Where can I download CVSS-X and its code?
The dataset is on Hugging Face at huggingface.co/datasets/lgris/XVSS-X under CC-BY-NC 4.0, and the code is on GitHub at github.com/ErmisAI/XVSS-X. Note that regenerating the corpus requires OmniVoice TTS and ElevenLabs voice design, which are not fully open.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The extraction includes PDF line-break artifacts mid-sentence (e.g., 'This recovered Common Voice English (v17) / 240,192 of the original...'); these are rendering breaks, not truncation — the full sentence continues in the text.
- Figure 1 is available only as its embedded text labels; no visual detail beyond the pipeline text was verifiable.
- The OpenReview and repository URLs are quoted as printed in the text and were not accessed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.