Hugging Face's logo

Datasets:
lgris
/
CVSS-X

Tasks:
Audio-to-Audio
Text-to-Speech
Translation
Languages:
English
Portuguese
Spanish
Size:
10M<n<100M
ArXiv:
Tags:
arxiv:2609.13413
speech-to-speech-translation
speech-translation
common-voice
cross-lingual-tts
voice-cloning
License:
Dataset card Data Studio Files
xet
 Community
CVSS-X / README.md
lgris's picture
lgris
Fix dataset name (XVSS-X -> CVSS-X)
a7e04be verified
preview code
|
Raw History Blame Contribute Delete
36.1 kB
---
configs:
- config_name: en-pt
data_files:
- split: train
path: data/en-pt/train/*/*.tar
- split: dev
path: data/en-pt/dev/*/*.tar
- split: test
path: data/en-pt/test/*/*.tar
- config_name: en-pt-canonical
data_files:
- split: train
path: data/en-pt/train/canonical/*.tar
- split: dev
path: data/en-pt/dev/canonical/*.tar
- split: test
path: data/en-pt/test/canonical/*.tar
- config_name: en-pt-timbre
data_files:
- split: train
path: data/en-pt/train/timbre/*.tar
- split: dev
path: data/en-pt/dev/timbre/*.tar
- split: test
path: data/en-pt/test/timbre/*.tar
- config_name: en-es
data_files:
- split: train
path: data/en-es/train/*/*.tar
- split: dev
path: data/en-es/dev/*/*.tar
- split: test
path: data/en-es/test/*/*.tar
- config_name: en-es-canonical
data_files:
- split: train
path: data/en-es/train/canonical/*.tar
- split: dev
path: data/en-es/dev/canonical/*.tar
- split: test
path: data/en-es/test/canonical/*.tar
- config_name: en-es-timbre
data_files:
- split: train
path: data/en-es/train/timbre/*.tar
- split: dev
path: data/en-es/dev/timbre/*.tar
- split: test
path: data/en-es/test/timbre/*.tar
- config_name: en-fr
data_files:
- split: train
path: data/en-fr/train/*/*.tar
- split: dev
path: data/en-fr/dev/*/*.tar
- split: test
path: data/en-fr/test/*/*.tar
- config_name: en-fr-canonical
data_files:
- split: train
path: data/en-fr/train/canonical/*.tar
- split: dev
path: data/en-fr/dev/canonical/*.tar
- split: test
path: data/en-fr/test/canonical/*.tar
- config_name: en-fr-timbre
data_files:
- split: train
path: data/en-fr/train/timbre/*.tar
- split: dev
path: data/en-fr/dev/timbre/*.tar
- split: test
path: data/en-fr/test/timbre/*.tar
- config_name: en-it
data_files:
- split: train
path: data/en-it/train/*/*.tar
- split: dev
path: data/en-it/dev/*/*.tar
- split: test
path: data/en-it/test/*/*.tar
- config_name: en-it-canonical
data_files:
- split: train
path: data/en-it/train/canonical/*.tar
- split: dev
path: data/en-it/dev/canonical/*.tar
- split: test
path: data/en-it/test/canonical/*.tar
- config_name: en-it-timbre
data_files:
- split: train
path: data/en-it/train/timbre/*.tar
- split: dev
path: data/en-it/dev/timbre/*.tar
- split: test
path: data/en-it/test/timbre/*.tar
- config_name: en-ro
data_files:
- split: train
path: data/en-ro/train/*/*.tar
- split: dev
path: data/en-ro/dev/*/*.tar
- split: test
path: data/en-ro/test/*/*.tar
- config_name: en-ro-canonical
data_files:
- split: train
path: data/en-ro/train/canonical/*.tar
- split: dev
path: data/en-ro/dev/canonical/*.tar
- split: test
path: data/en-ro/test/canonical/*.tar
- config_name: en-ro-timbre
data_files:
- split: train
path: data/en-ro/train/timbre/*.tar
- split: dev
path: data/en-ro/dev/timbre/*.tar
- split: test
path: data/en-ro/test/timbre/*.tar
- config_name: en-ca
data_files:
- split: train
path: data/en-ca/train/*/*.tar
- split: dev
path: data/en-ca/dev/*/*.tar
- split: test
path: data/en-ca/test/*/*.tar
- config_name: en-ca-canonical
data_files:
- split: train
path: data/en-ca/train/canonical/*.tar
- split: dev
path: data/en-ca/dev/canonical/*.tar
- split: test
path: data/en-ca/test/canonical/*.tar
- config_name: en-ca-timbre
data_files:
- split: train
path: data/en-ca/train/timbre/*.tar
- split: dev
path: data/en-ca/dev/timbre/*.tar
- split: test
path: data/en-ca/test/timbre/*.tar
- config_name: en-de
data_files:
- split: train
path: data/en-de/train/*/*.tar
- split: dev
path: data/en-de/dev/*/*.tar
- split: test
path: data/en-de/test/*/*.tar
- config_name: en-de-canonical
data_files:
- split: train
path: data/en-de/train/canonical/*.tar
- split: dev
path: data/en-de/dev/canonical/*.tar
- split: test
path: data/en-de/test/canonical/*.tar
- config_name: en-de-timbre
data_files:
- split: train
path: data/en-de/train/timbre/*.tar
- split: dev
path: data/en-de/dev/timbre/*.tar
- split: test
path: data/en-de/test/timbre/*.tar
- config_name: en-nl
data_files:
- split: train
path: data/en-nl/train/*/*.tar
- split: dev
path: data/en-nl/dev/*/*.tar
- split: test
path: data/en-nl/test/*/*.tar
- config_name: en-nl-canonical
data_files:
- split: train
path: data/en-nl/train/canonical/*.tar
- split: dev
path: data/en-nl/dev/canonical/*.tar
- split: test
path: data/en-nl/test/canonical/*.tar
- config_name: en-nl-timbre
data_files:
- split: train
path: data/en-nl/train/timbre/*.tar
- split: dev
path: data/en-nl/dev/timbre/*.tar
- split: test
path: data/en-nl/test/timbre/*.tar
- config_name: en-sv
data_files:
- split: train
path: data/en-sv/train/*/*.tar
- split: dev
path: data/en-sv/dev/*/*.tar
- split: test
path: data/en-sv/test/*/*.tar
- config_name: en-sv-canonical
data_files:
- split: train
path: data/en-sv/train/canonical/*.tar
- split: dev
path: data/en-sv/dev/canonical/*.tar
- split: test
path: data/en-sv/test/canonical/*.tar
- config_name: en-sv-timbre
data_files:
- split: train
path: data/en-sv/train/timbre/*.tar
- split: dev
path: data/en-sv/dev/timbre/*.tar
- split: test
path: data/en-sv/test/timbre/*.tar
- config_name: en-da
data_files:
- split: train
path: data/en-da/train/*/*.tar
- split: dev
path: data/en-da/dev/*/*.tar
- split: test
path: data/en-da/test/*/*.tar
- config_name: en-da-canonical
data_files:
- split: train
path: data/en-da/train/canonical/*.tar
- split: dev
path: data/en-da/dev/canonical/*.tar
- split: test
path: data/en-da/test/canonical/*.tar
- config_name: en-da-timbre
data_files:
- split: train
path: data/en-da/train/timbre/*.tar
- split: dev
path: data/en-da/dev/timbre/*.tar
- split: test
path: data/en-da/test/timbre/*.tar
- config_name: en-no
data_files:
- split: train
path: data/en-no/train/*/*.tar
- split: dev
path: data/en-no/dev/*/*.tar
- split: test
path: data/en-no/test/*/*.tar
- config_name: en-no-canonical
data_files:
- split: train
path: data/en-no/train/canonical/*.tar
- split: dev
path: data/en-no/dev/canonical/*.tar
- split: test
path: data/en-no/test/canonical/*.tar
- config_name: en-no-timbre
data_files:
- split: train
path: data/en-no/train/timbre/*.tar
- split: dev
path: data/en-no/dev/timbre/*.tar
- split: test
path: data/en-no/test/timbre/*.tar
- config_name: en-ru
data_files:
- split: train
path: data/en-ru/train/*/*.tar
- split: dev
path: data/en-ru/dev/*/*.tar
- split: test
path: data/en-ru/test/*/*.tar
- config_name: en-ru-canonical
data_files:
- split: train
path: data/en-ru/train/canonical/*.tar
- split: dev
path: data/en-ru/dev/canonical/*.tar
- split: test
path: data/en-ru/test/canonical/*.tar
- config_name: en-ru-timbre
data_files:
- split: train
path: data/en-ru/train/timbre/*.tar
- split: dev
path: data/en-ru/dev/timbre/*.tar
- split: test
path: data/en-ru/test/timbre/*.tar
- config_name: en-pl
data_files:
- split: train
path: data/en-pl/train/*/*.tar
- split: dev
path: data/en-pl/dev/*/*.tar
- split: test
path: data/en-pl/test/*/*.tar
- config_name: en-pl-canonical
data_files:
- split: train
path: data/en-pl/train/canonical/*.tar
- split: dev
path: data/en-pl/dev/canonical/*.tar
- split: test
path: data/en-pl/test/canonical/*.tar
- config_name: en-pl-timbre
data_files:
- split: train
path: data/en-pl/train/timbre/*.tar
- split: dev
path: data/en-pl/dev/timbre/*.tar
- split: test
path: data/en-pl/test/timbre/*.tar
- config_name: en-cs
data_files:
- split: train
path: data/en-cs/train/*/*.tar
- split: dev
path: data/en-cs/dev/*/*.tar
- split: test
path: data/en-cs/test/*/*.tar
- config_name: en-cs-canonical
data_files:
- split: train
path: data/en-cs/train/canonical/*.tar
- split: dev
path: data/en-cs/dev/canonical/*.tar
- split: test
path: data/en-cs/test/canonical/*.tar
- config_name: en-cs-timbre
data_files:
- split: train
path: data/en-cs/train/timbre/*.tar
- split: dev
path: data/en-cs/dev/timbre/*.tar
- split: test
path: data/en-cs/test/timbre/*.tar
- config_name: en-uk
data_files:
- split: train
path: data/en-uk/train/*/*.tar
- split: dev
path: data/en-uk/dev/*/*.tar
- split: test
path: data/en-uk/test/*/*.tar
- config_name: en-uk-canonical
data_files:
- split: train
path: data/en-uk/train/canonical/*.tar
- split: dev
path: data/en-uk/dev/canonical/*.tar
- split: test
path: data/en-uk/test/canonical/*.tar
- config_name: en-uk-timbre
data_files:
- split: train
path: data/en-uk/train/timbre/*.tar
- split: dev
path: data/en-uk/dev/timbre/*.tar
- split: test
path: data/en-uk/test/timbre/*.tar
- config_name: en-zh
data_files:
- split: train
path: data/en-zh/train/*/*.tar
- split: dev
path: data/en-zh/dev/*/*.tar
- split: test
path: data/en-zh/test/*/*.tar
- config_name: en-zh-canonical
data_files:
- split: train
path: data/en-zh/train/canonical/*.tar
- split: dev
path: data/en-zh/dev/canonical/*.tar
- split: test
path: data/en-zh/test/canonical/*.tar
- config_name: en-zh-timbre
data_files:
- split: train
path: data/en-zh/train/timbre/*.tar
- split: dev
path: data/en-zh/dev/timbre/*.tar
- split: test
path: data/en-zh/test/timbre/*.tar
- config_name: en-ja
data_files:
- split: train
path: data/en-ja/train/*/*.tar
- split: dev
path: data/en-ja/dev/*/*.tar
- split: test
path: data/en-ja/test/*/*.tar
- config_name: en-ja-canonical
data_files:
- split: train
path: data/en-ja/train/canonical/*.tar
- split: dev
path: data/en-ja/dev/canonical/*.tar
- split: test
path: data/en-ja/test/canonical/*.tar
- config_name: en-ja-timbre
data_files:
- split: train
path: data/en-ja/train/timbre/*.tar
- split: dev
path: data/en-ja/dev/timbre/*.tar
- split: test
path: data/en-ja/test/timbre/*.tar
- config_name: en-ko
data_files:
- split: train
path: data/en-ko/train/*/*.tar
- split: dev
path: data/en-ko/dev/*/*.tar
- split: test
path: data/en-ko/test/*/*.tar
- config_name: en-ko-canonical
data_files:
- split: train
path: data/en-ko/train/canonical/*.tar
- split: dev
path: data/en-ko/dev/canonical/*.tar
- split: test
path: data/en-ko/test/canonical/*.tar
- config_name: en-ko-timbre
data_files:
- split: train
path: data/en-ko/train/timbre/*.tar
- split: dev
path: data/en-ko/dev/timbre/*.tar
- split: test
path: data/en-ko/test/timbre/*.tar
- config_name: en-fi
data_files:
- split: train
path: data/en-fi/train/*/*.tar
- split: dev
path: data/en-fi/dev/*/*.tar
- split: test
path: data/en-fi/test/*/*.tar
- config_name: en-fi-canonical
data_files:
- split: train
path: data/en-fi/train/canonical/*.tar
- split: dev
path: data/en-fi/dev/canonical/*.tar
- split: test
path: data/en-fi/test/canonical/*.tar
- config_name: en-fi-timbre
data_files:
- split: train
path: data/en-fi/train/timbre/*.tar
- split: dev
path: data/en-fi/dev/timbre/*.tar
- split: test
path: data/en-fi/test/timbre/*.tar
- config_name: en-hu
data_files:
- split: train
path: data/en-hu/train/*/*.tar
- split: dev
path: data/en-hu/dev/*/*.tar
- split: test
path: data/en-hu/test/*/*.tar
- config_name: en-hu-canonical
data_files:
- split: train
path: data/en-hu/train/canonical/*.tar
- split: dev
path: data/en-hu/dev/canonical/*.tar
- split: test
path: data/en-hu/test/canonical/*.tar
- config_name: en-hu-timbre
data_files:
- split: train
path: data/en-hu/train/timbre/*.tar
- split: dev
path: data/en-hu/dev/timbre/*.tar
- split: test
path: data/en-hu/test/timbre/*.tar
- config_name: en-hi
data_files:
- split: train
path: data/en-hi/train/*/*.tar
- split: dev
path: data/en-hi/dev/*/*.tar
- split: test
path: data/en-hi/test/*/*.tar
- config_name: en-hi-canonical
data_files:
- split: train
path: data/en-hi/train/canonical/*.tar
- split: dev
path: data/en-hi/dev/canonical/*.tar
- split: test
path: data/en-hi/test/canonical/*.tar
- config_name: en-hi-timbre
data_files:
- split: train
path: data/en-hi/train/timbre/*.tar
- split: dev
path: data/en-hi/dev/timbre/*.tar
- split: test
path: data/en-hi/test/timbre/*.tar
- config_name: en-fa
data_files:
- split: train
path: data/en-fa/train/*/*.tar
- split: dev
path: data/en-fa/dev/*/*.tar
- split: test
path: data/en-fa/test/*/*.tar
- config_name: en-fa-canonical
data_files:
- split: train
path: data/en-fa/train/canonical/*.tar
- split: dev
path: data/en-fa/dev/canonical/*.tar
- split: test
path: data/en-fa/test/canonical/*.tar
- config_name: en-fa-timbre
data_files:
- split: train
path: data/en-fa/train/timbre/*.tar
- split: dev
path: data/en-fa/dev/timbre/*.tar
- split: test
path: data/en-fa/test/timbre/*.tar
- config_name: en-el
data_files:
- split: train
path: data/en-el/train/*/*.tar
- split: dev
path: data/en-el/dev/*/*.tar
- split: test
path: data/en-el/test/*/*.tar
- config_name: en-el-canonical
data_files:
- split: train
path: data/en-el/train/canonical/*.tar
- split: dev
path: data/en-el/dev/canonical/*.tar
- split: test
path: data/en-el/test/canonical/*.tar
- config_name: en-el-timbre
data_files:
- split: train
path: data/en-el/train/timbre/*.tar
- split: dev
path: data/en-el/dev/timbre/*.tar
- split: test
path: data/en-el/test/timbre/*.tar
- config_name: en-he
data_files:
- split: train
path: data/en-he/train/*/*.tar
- split: dev
path: data/en-he/dev/*/*.tar
- split: test
path: data/en-he/test/*/*.tar
- config_name: en-he-canonical
data_files:
- split: train
path: data/en-he/train/canonical/*.tar
- split: dev
path: data/en-he/dev/canonical/*.tar
- split: test
path: data/en-he/test/canonical/*.tar
- config_name: en-he-timbre
data_files:
- split: train
path: data/en-he/train/timbre/*.tar
- split: dev
path: data/en-he/dev/timbre/*.tar
- split: test
path: data/en-he/test/timbre/*.tar
- config_name: en-tr
data_files:
- split: train
path: data/en-tr/train/*/*.tar
- split: dev
path: data/en-tr/dev/*/*.tar
- split: test
path: data/en-tr/test/*/*.tar
- config_name: en-tr-canonical
data_files:
- split: train
path: data/en-tr/train/canonical/*.tar
- split: dev
path: data/en-tr/dev/canonical/*.tar
- split: test
path: data/en-tr/test/canonical/*.tar
- config_name: en-tr-timbre
data_files:
- split: train
path: data/en-tr/train/timbre/*.tar
- split: dev
path: data/en-tr/dev/timbre/*.tar
- split: test
path: data/en-tr/test/timbre/*.tar
- config_name: en-th
data_files:
- split: train
path: data/en-th/train/*/*.tar
- split: dev
path: data/en-th/dev/*/*.tar
- split: test
path: data/en-th/test/*/*.tar
- config_name: en-th-canonical
data_files:
- split: train
path: data/en-th/train/canonical/*.tar
- split: dev
path: data/en-th/dev/canonical/*.tar
- split: test
path: data/en-th/test/canonical/*.tar
- config_name: en-th-timbre
data_files:
- split: train
path: data/en-th/train/timbre/*.tar
- split: dev
path: data/en-th/dev/timbre/*.tar
- split: test
path: data/en-th/test/timbre/*.tar
- config_name: en-id
data_files:
- split: train
path: data/en-id/train/*/*.tar
- split: dev
path: data/en-id/dev/*/*.tar
- split: test
path: data/en-id/test/*/*.tar
- config_name: en-id-canonical
data_files:
- split: train
path: data/en-id/train/canonical/*.tar
- split: dev
path: data/en-id/dev/canonical/*.tar
- split: test
path: data/en-id/test/canonical/*.tar
- config_name: en-id-timbre
data_files:
- split: train
path: data/en-id/train/timbre/*.tar
- split: dev
path: data/en-id/dev/timbre/*.tar
- split: test
path: data/en-id/test/timbre/*.tar
- config_name: en-vi
data_files:
- split: train
path: data/en-vi/train/*/*.tar
- split: dev
path: data/en-vi/dev/*/*.tar
- split: test
path: data/en-vi/test/*/*.tar
- config_name: en-vi-canonical
data_files:
- split: train
path: data/en-vi/train/canonical/*.tar
- split: dev
path: data/en-vi/dev/canonical/*.tar
- split: test
path: data/en-vi/test/canonical/*.tar
- config_name: en-vi-timbre
data_files:
- split: train
path: data/en-vi/train/timbre/*.tar
- split: dev
path: data/en-vi/dev/timbre/*.tar
- split: test
path: data/en-vi/test/timbre/*.tar
task_categories:
- audio-to-audio
- text-to-speech
- translation
language:
- en
- pt
- es
- fr
- it
- ro
- ca
- de
- nl
- sv
- da
- no
- ru
- pl
- cs
- uk
- zh
- ja
- ko
- fi
- hu
- hi
- fa
- el
- he
- tr
- th
- id
- vi
license: cc-by-nc-4.0
tags:
- arxiv:2609.13413
- speech-to-speech-translation
- speech-translation
- common-voice
- cross-lingual-tts
- voice-cloning
size_categories:
- 10M<n<100M
---
# CVSS-X: A Multilingual Speech-to-Speech Translation Corpus for 28 Languages
[![License: CC BY-NC 4.0](https://img.shields.io/badge/License-CC%20BY--NC%204.0-lightgrey.svg)](https://creativecommons.org/licenses/by-nc/4.0/)
[![arXiv](https://img.shields.io/badge/arXiv-2609.13413-b31b1b.svg)](https://arxiv.org/abs/2609.13413)
[![Python 3.10+](https://img.shields.io/badge/Python-3.10%2B-blue.svg)](https://www.python.org/)
[![PyTorch 2.1+](https://img.shields.io/badge/PyTorch-2.1%2B-orange.svg)](https://pytorch.org/)
[![Paper](https://img.shields.io/badge/Paper-Accepted-green.svg)](#citation)
This repository contains the official, reproducible pipeline and evaluation suite for **CVSS-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**).
---
## Table of Contents
- [Overview](#overview)
- [Corpus Statistics](#corpus-statistics)
- [Target Languages](#target-languages)
- [Dataset Generation Pipeline](#dataset-generation-pipeline)
- [Quality Evaluation Results](#quality-evaluation-results)
- [Repository Structure](#repository-structure)
- [Quickstart & Reproduction](#quickstart--reproduction)
- [1. Environment Setup](#1-environment-setup)
- [2. Step 1: Source Preparation (Official vs. Custom)](#2-step-1-source-preparation-official-vs-custom)
- [3. Step 2: Multilingual Translation](#3-step-2-multilingual-translation)
- [4. Step 3 & 4: Speech Synthesis](#4-step-3--4-speech-synthesis)
- [5. Step 5: Evaluation Suite](#5-step-5-evaluation-suite)
- [Hugging Face Dataset Access](#hugging-face-dataset-access)
- [Roadmap: Version 2 (In Development)](#roadmap-version-2-in-development)
- [License](#license)
- [Citation](#citation)
---
## Overview
While the pioneering CVSS corpus enables translation from 21 source languages exclusively into English (many-to-one), **CVSS-X** provides one-to-many translation from English into **28 typologically diverse target languages** across 12 language families.
Combined with CVSS, CVSS-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.
CVSS-X is provided in two complementary variants:
- **CVSS-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).
- **CVSS-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](https://github.com/k2-fsa/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** |
| **CVSS-X-C Duration** | ~6,730 hours (avg. 3.6s / utterance) |
| **CVSS-X-T Duration** | ~9,340 hours (avg. 5.0s / utterance) |
| **Total Audio Duration** | **~16,070 hours** (8x larger than CVSS) |
---
## Target Languages
CVSS-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)
└─────────────┬─────────────┘
│
┌───────┴───────────────────────┐
▼ ▼
CVSS-X-C (Canonical) CVSS-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 | CVSS-X-C | CVSS-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
```
CVSS-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:
```bash
git clone https://github.com/Ermisai/CVSS-X.git
cd CVSS-X
./setup_env.sh
conda activate cvss-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/`:
```bash
# 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:
```bash
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):
```bash
# 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:
```bash
# 1. Canonical Variant (CVSS-X-C):
./scripts/03_run_synthesis_canonical.sh
# 2. Timbre-Transferred Variant (CVSS-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):
```bash
./scripts/05_run_evaluation.sh
```
The resulting Markdown report and paper tables will be generated in `evaluation/evaluation_report.md`.
---
## Hugging Face Dataset Access & WebDataset Streaming
The complete dataset (~2.7 TB across 28 language pairs, 13.4M+ audio files) is being mirrored to the Hugging Face Hub under the repository [`lgris/CVSS-X`](https://huggingface.co/datasets/lgris/CVSS-X).
### 1. Standard Streaming via `load_dataset`
You can stream subsets directly per language pair without downloading the entire corpus:
```python
from datasets import load_dataset
# Stream Portuguese canonical dev split
dataset = load_dataset("lgris/CVSS-X", "en-pt", split="dev", streaming=True)
sample = next(iter(dataset))
print(sample.keys())
```
### 2. Loading via WebDataset Format
Audio files are packaged into sequential `.tar` archives natively compatible with the [Hugging Face Datasets WebDataset loader](https://huggingface.co/docs/datasets/en/loading#webdataset). You can stream shards directly:
```python
from datasets import load_dataset
# Load WebDataset shards directly in streaming mode
dataset_wds = load_dataset(
"webdataset",
data_files={"train": "https://huggingface.co/datasets/lgris/CVSS-X/resolve/main/data/en-pt/train/*/*.tar"},
split="train",
streaming=True
)
sample_wds = next(iter(dataset_wds))
# WebDataset automatically decodes the audio feature
audio = sample_wds["wav"]
print(f"Sample key: {sample_wds['__key__']}, audio sampling rate: {audio['sampling_rate']}")
```
---
## Roadmap: Version 2 & Downstream Models (In Development)
> [!NOTE]
> **We are actively working on Version 2 (v2) of the dataset and downstream baseline models.**
Key improvements and models currently planned or in progress:
- **TAGARELA v2 Baseline Model Integration**: Adapting and training discrete-token speech-to-speech translation baselines directly on CVSS-X using the [TAGARELA v2](https://github.com/freds0/TAGARELA_v2) architecture ([`freds0/TAGARELA_v2`](https://github.com/freds0/TAGARELA_v2)).
- **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.
---
## License
The CVSS-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](LICENSE) for full legal text.
---
## Citation
If you use CVSS-X in your research, please cite our paper:
```bibtex
@misc{gris2026cvssx,
title={{CVSS-X: A Multilingual Speech-to-Speech Translation Corpus for 28 Languages}},
author={Lucas Rafael Stefanel Gris and Alef Iury Siqueira Ferreira and Frederico Santos de Oliveira and Augusto Seben da Rosa and Alexandre Costa Ferro Filho and Arlindo Rodrigues Galv{\~a}o Filho and Anderson da Silva Soares},
year={2026},
eprint={2609.13413},
archivePrefix={arXiv},
primaryClass={cs.CL},
doi={10.48550/arXiv.2609.13413},
url={https://arxiv.org/abs/2609.13413}
}
```
Preprint available on arXiv: [https://arxiv.org/abs/2609.13413](https://arxiv.org/abs/2609.13413)