XVSS-X: A Multilingual Speech-to-Speech Translation Corpus for 28 Languages

This repository contains the official, reproducible pipeline and evaluation suite for XVSS-X (CVSS-X), a massive synthetic speech-to-speech translation (S2ST) corpus that complements and extends the original CVSS dataset by reversing the translation direction (English → 28 target languages).

---

---

Overview

While the pioneering CVSS corpus enables translation from 21 source languages exclusively into English (many-to-one), XVSS-X provides one-to-many translation from English into 28 typologically diverse target languages across 12 language families.

Combined with CVSS, XVSS-X enables:

1. Bidirectional speech-to-speech translation (English ↔ 28 languages).
2. Multilingual translation between arbitrary language pairs using English as a pivot (X → EN → Y).
3. Direct evaluation of cross-lingual voice preservation in speech-to-speech tasks.

XVSS-X is provided in two complementary variants:

• XVSS-X-C (Canonical): Target speech is synthesized using two fixed, high-clarity reference voices per language (one male, one female). Voice selection is conditioned on the source speaker's gender metadata (81.4% male, 18.6% female).
• XVSS-X-T (Timbre-Transferred): Preserves the speaker characteristics of the original English speaker in the target speech via zero-shot cross-lingual voice cloning with OmniVoice.

---

Corpus Statistics

Metric	Value
Source Language	English (Common Voice 17)
Target Languages	28
Language Families	12 (7 macro-families)
Nominal Samples per Language	240,192
Data Splits (per lang)	Train: 222,349 • Dev: 10,000 • Test: 7,843
Total Parallel Speech Pairs	6,725,176 (Canonical) + 6,724,326 (Timbre) = 13,449,502 audio files
XVSS-X-C Duration	~6,730 hours (avg. 3.6s / utterance)
XVSS-X-T Duration	~9,340 hours (avg. 5.0s / utterance)
Total Audio Duration	~16,070 hours (8x larger than CVSS)

---

Target Languages

XVSS-X spans 28 languages across 12 typological families:

Family	N	Languages (ISO Codes)
Romance	6	Portuguese (pt), Spanish (es), French (fr), Italian (it), Romanian (ro), Catalan (ca)
Germanic	5	German (de), Dutch (nl), Swedish (sv), Danish (da), Norwegian (no)
Slavic	4	Russian (ru), Polish (pl), Czech (cs), Ukrainian (uk)
CJK	3	Chinese (zh), Japanese (ja), Korean (ko)
Uralic	2	Finnish (fi), Hungarian (hu)
Indo-Iranian	2	Hindi (hi), Persian (fa)
Other	6	Greek (el), Hebrew (he), Turkish (tr), Thai (th), Indonesian (id), Vietnamese (vi)

---

Dataset Generation Pipeline

Common Voice 17 EN
  (240K utterances)
        │
        ▼
┌───────────────────────────┐
│   1. Source Alignment     │ ──► Aligns CV17 recordings with CVSS metadata
└─────────────┬─────────────┘     (train: 222,349 | dev: 10,000 | test: 7,843)
              │
              ▼
┌───────────────────────────┐
│   2. Text Translation     │ ──► facebook/nllb-200-distilled-600M
└─────────────┬─────────────┘     (EN -> 28 target languages, ~17ms/sentence)
              │
              ▼
┌───────────────────────────┐
│   3. Speech Synthesis     │ ──► k2-fsa/OmniVoice (zero-shot cross-lingual TTS)
└─────────────┬─────────────┘
              │
      ┌───────┴───────────────────────┐
      ▼                               ▼
XVSS-X-C (Canonical)           XVSS-X-T (Timbre-Transferred)
2 fixed reference voices       Cloned voice from English source
(~6,730 hours)                 (~9,340 hours)

1. Source Data & Alignment: English recordings from Common Voice version 17 aligned via normalized text matching with the original CVSS corpus (recovering 91.0% of samples). The dev set is supplemented from the training split to reach exactly 10,000 samples.
2. Text Translation: Transcripts translated using facebook/nllb-200-distilled-600M, selected after extensive benchmarking across 7 translation models for its optimal quality/throughput trade-off.
3. Speech Synthesis: Generated with k2-fsa/OmniVoice, a multilingual transformer TTS engine featuring zero-shot cross-lingual synthesis without transferring source accent artifacts.

---

Quality Evaluation Results

Evaluation performed on a stratified random sample of 200 utterances per language from the dev set (5,600 samples per variant, determined via statistical power analysis).

Evaluation by Language Family (CVSS-X)

Family (N)	WER/CER (C)	WER/CER (T)	ASR-BLEU (C)	ASR-BLEU (T)	UTMOS (C)	UTMOS (T)
Romance (6)	5.8	8.2	90.3	88.0	3.54	3.27
Germanic (5)	9.5	12.2	85.3	81.1	3.62	3.27
Slavic (4)	7.0	7.7	86.9	85.9	3.48	3.17
CJK (3)*	5.0	10.2	75.2	67.9	3.56	3.20
Uralic (2)	10.3	14.3	84.0	78.3	3.54	3.21
Indo-Iranian (2)	22.5	23.0	63.7	60.0	3.61	3.20
Other (6)	24.8	24.8	65.0	64.9	3.53	3.14
Average	12.1	14.1	82.4	79.4	3.55	3.21

*For unsegmented CJK languages (ZH, JA, KO), Character Error Rate (CER) and character-level BLEU are reported.

Overall Comparison with CVSS (Re-evaluated with Same Pipeline)

Metric	XVSS-X-C	XVSS-X-T	CVSS-C	CVSS-T
UTMOS (1–5)	3.55	3.21	4.43	3.61
ASR-BLEU	82.4	79.4	94.2	93.8
WER/CER (%)	12.1	14.1	3.5	4.0
Speaker Similarity (ECAPA-TDNN)	--	0.607	--	--

---

Repository Structure

XVSS-X/
├── README.md                      # Dataset documentation and reproduction guide
├── LICENSE                        # CC-BY-NC 4.0 license
├── requirements.txt               # Unified Python dependencies
├── setup_env.sh                   # One-stop environment setup script
├── .gitignore                     # Git ignore rules
│
├── metadata/                      # Official Pre-computed Manifests
│   ├── manifest_train.json        # 222,349 source utterances
│   ├── manifest_dev.json          # 10,000 source utterances
│   └── manifest_test.json         # 7,843 source utterances
│
├── config/
│   ├── languages.yaml             # Complete metadata for all 28 target languages
│   └── pipeline.yaml              # Default hyperparameters and paths
│
├── pipeline/                      # Dataset Creation Pipeline (Clean & Modular)
│   ├── __init__.py
│   ├── 01_prepare_source.py       # CV17 extraction, CVSS alignment, and split generation
│   ├── 02_translate.py            # NLLB multilingual translation (EN -> 28 languages)
│   ├── 03_synthesize.py           # OmniVoice TTS synthesis (Canonical & Timbre modes)
│   └── utils.py                   # Data schemas, manifest I/O, checkpoints, audio helpers
│
├── evaluation/                    # Quality Evaluation Suite
│   ├── __init__.py
│   ├── sample_dev.py              # Deterministic 200 utterances/lang dev sampling (seed=42)
│   ├── evaluate_asr_bleu.py       # Whisper large-v3 ASR + SacreBLEU / chrF / CER / WER
│   ├── evaluate_utmos.py          # Neural MOS speech naturalness (UTMOS)
│   ├── evaluate_speaker_sim.py    # ECAPA-TDNN cross-lingual speaker similarity
│   ├── evaluate_cvss_baseline.py  # Re-evaluation of original CVSS baseline
│   └── generate_tables.py         # Summary report & LaTeX/Markdown table generator
│
├── scripts/                       # Reproducible Bash Runners
│   ├── 01_run_data_prep.sh        # Executes Step 1 (source data prep or load official)
│   ├── 02_run_translation.sh      # Executes Step 2 (translation)
│   ├── 03_run_synthesis_canonical.sh # Executes Step 3 (canonical synthesis)
│   ├── 04_run_synthesis_timbre.sh    # Executes Step 4 (timbre synthesis)
│   └── 05_run_evaluation.sh       # Executes Step 5 (full evaluation suite)
│
└── assets/                        # Reference voice templates
    ├── pt-male.wav
    └── pt-female.wav

---

Quickstart & Reproduction

1. Environment Setup

Clone this repository and run the setup script:

git clone https://github.com/ErmisAI/XVSS-X.git
cd XVSS-X
./setup_env.sh
conda activate xvss-x  # or source .venv/bin/activate

2. Step 1: Source Preparation (Official vs. Custom)

You have two options to prepare the source data:

Option A: Use the Official Pre-computed Manifests (Recommended)

You can directly use the exact aligned splits included in this repository under metadata/:

# Export official manifests directly to data/manifests/
USE_OFFICIAL_MANIFESTS=true ./scripts/01_run_data_prep.sh

# Or optionally link your local Common Voice 17 audio directory:
USE_OFFICIAL_MANIFESTS=true CV17_DIR=/path/to/cv17 ./scripts/01_run_data_prep.sh

Option B: Recompute Alignment from Scratch

If you wish to re-execute text alignment from raw Common Voice v4 and v17:

CV4_DIR=/path/to/cv4 CV17_DIR=/path/to/cv17 ./scripts/01_run_data_prep.sh

This creates data/manifests/manifest_{train,dev,test}.json containing exactly 222,349 / 10,000 / 7,843 samples.

3. Step 2: Multilingual Translation

Translate English transcripts to all 28 target languages (or a specific subset):

# Translate all 28 languages:
./scripts/02_run_translation.sh

# Or translate specific languages (e.g., Portuguese and Spanish):
LANGS="pt,es" ./scripts/02_run_translation.sh

Checkpoints are saved automatically every 2,000 samples for seamless resume capability.

4. Step 3 & 4: Speech Synthesis

Generate speech using OmniVoice:

# 1. Canonical Variant (XVSS-X-C):
./scripts/03_run_synthesis_canonical.sh

# 2. Timbre-Transferred Variant (XVSS-X-T):
./scripts/04_run_synthesis_timbre.sh

Audio files are saved under data/synthesized/en-{lang}/{split}/{variant}/{sample_id}.wav at 24kHz.

5. Step 5: Evaluation Suite

Run the full evaluation pipeline (dev sampling, ASR-BLEU with Whisper large-v3, UTMOS, and ECAPA-TDNN):

./scripts/05_run_evaluation.sh

The resulting Markdown report and paper tables will be generated in evaluation/evaluation_report.md.

---

Hugging Face Dataset Access

The complete dataset (~2.7 TB across 28 language pairs, 13.4M+ audio files) is being mirrored to the Hugging Face Hub under the account lgris/XVSS-X.

You will be able to load and stream subsets directly using the datasets library:

from datasets import load_dataset

# Stream Portuguese canonical dev split
dataset = load_dataset("lgris/XVSS-X", "en-pt", split="dev", streaming=True)
sample = next(iter(dataset))
print(sample["source_text"])
print(sample["target_text"])

---

Roadmap: Version 2 (In Development)

We are actively working on Version 2 (v2) of the dataset.

Key improvements planned for XVSS-X v2 include:

• TranslateGemma-12B Integration: Replacing NLLB-200 with Google's TranslateGemma-12B to significantly improve translation fidelity, natural phrasing, and eliminate commercial restrictions by transitioning to an Apache 2.0 license.
• Common Voice 26 & Spontaneous Speech 4.0: Expanding from CV17 to Common Voice 26 and integrating Spontaneous Speech 4.0, offering far richer conversational speech, more realistic disfluencies, and greater speaker diversity.
• End-to-End LLM Baselines: Training and evaluating native discrete audio token speech-to-speech translation baselines.

---

License

The XVSS-X corpus and code are licensed as follows:

• Audio & Source Transcripts: Inherited from Mozilla Common Voice (released under CC0 1.0 Universal).
• Synthetic Translations & Speech Pairs: Distributed under Creative Commons Attribution-NonCommercial 4.0 International (CC-BY-NC 4.0) due to the underlying facebook/nllb-200-distilled-600M model license.
• Codebase & Scripts: Released under the permissive MIT License.

See LICENSE for full legal text.

---

Citation

If you use XVSS-X in your research, please cite our paper:

@inproceedings{gris2026cvssx,
  title={{CVSS-X: A Multilingual Speech-to-Speech Translation Corpus for 28 Languages}},
  author={Gris, Lucas Rafael Stefanel and Ferreira, Alef Iury Siqueira and de Oliveira, F. S. and da Rosa, Augusto Seben and Ferro Filho, Alexandre Costa and Galv{\~a}o Filho, Arlindo Rodrigues and Soares, Anderson da Silva},
  booktitle={Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL)},
  year={2026}
}