meetween/mumospee_covost

Dataset Summary

This dataset is a modified version of the CoVoST corpus, converted into parquet format to facilitate optimized I/O operations in high-performance and distributed computing environments. The CoVoST dataset is widely used in multilingual speech-to-text translation, containing over 2,800 hours of audio across 21 languages. This reformatted version retains the diverse multilingual and speaker attributes of CoVoST, with enhancements tailored for efficient data handling and training in large-scale machine learning workflows.

---

Source Data

• Original Dataset: CoVoST (built on top of Mozilla Common Voice).
• License: This derived dataset is shared under the same license, with modifications only to format for efficiency.

Modifications

• Data Format: Converted to parquet format to enhance I/O performance for distributed training, reducing latency during data loading and retrieval.
• Efficiency Optimization: Restructured for reduced storage footprint and faster I/O on high-performance clusters by leveraging parquet's efficient compression and columnar storage.
• Per-row duration: A duration column (seconds, computed once from the audio) is included in every data file, so downstream code does not need to decode audio just to filter by length.
• Companion index: A lightweight metadata/ folder mirrors each data parquet at the row level without the audio bytes — see "Dataset Structure" below.

Dataset Structure

• File Format: Parquet files.
• Coverage: Full CoVoST 2 — 36 language pairs across 21 source languages and 15 English-to-X targets:
  • English → X (15 pairs): ar, ca, cy, de, et, fa, id, ja, lv, mn, sl, sv-SE, ta, tr, zh-CN
  • X → English (21 pairs): ar, ca, cy, de, es, et, fa, fr, id, it, ja, lv, mn, nl, pt, ru, sl, sv-SE, ta, tr, zh-CN
• Each pair has train / validation / test splits, giving 108 data parquets in total.
• Audio Sampling Rate: Matches original dataset specifications for high-fidelity speech data.
• Multilingual Speaker Representation: Over 2,800 hours across multiple languages, preserving CoVoST's speaker diversity and multilingual alignment.

Repository layout

covost-parquet/
  en_xx/   covost2.en_<X>.<split>.parquet     ← audio + parallel text (the actual data)
  xx_en/   covost2.<X>_en.<split>.parquet
metadata/
           covost2.<pair>.<split>.parquet     ← per-row index (no audio bytes)

The two folders have the same number of files (108 each) and the same row counts and row order per file, but they serve different purposes — see below.

covost-parquet/ — the data (load this to train a model)

Schema (4 columns):

Column	Type	Description
audio	struct<bytes: binary, path: string>	MP3 bytes inline + the original Common Voice filename
duration	double	Audio length in seconds
transcription	string	Source-language text
translation	string	Target-language text

File sizes: ~13 MB (small validation sets) up to ~12 GB (en→X train sets, which share the full English Common Voice).

metadata/ — a row-level index (load this for cheap enumeration / multi-corpus indexing)

Each metadata parquet mirrors the rows of the corresponding covost-parquet/... file without the audio bytes and without the target-side translation. It is what you load when you want to enumerate, filter, or merge CoVoST entries without paying the cost of pulling audio. The schema is identical (column names, order, and types) to meetween/mumospee, so these files can be concatenated into a multi-corpus manifest.

Schema (9 columns, all string):

Column	Value
path	Original Common Voice filename (matches audio.path in the data parquet)
url	HF URL of the data parquet containing this row's audio
type	"audio"
duration	Audio length in seconds, serialized as a string (e.g. "5.165625")
language	Source-side language code (e.g. "en" for en_de, "de" for de_en)
transcript	Source-language text (same content as transcription in the data parquet)
tag	"CoVoST"
split	"train" / "validation" / "test"
license	"CC0"

Which folder do I use?

If you want to…	Load
Train an ASR or ST model on the audio + text	covost-parquet/...
Enumerate utterances, filter by duration / language, build a manifest	metadata/...
Merge CoVoST entries into a multi-corpus index	metadata/... (schema matches meetween/mumospee)
Look up the source-language transcript only	either (both have it)
Look up the target-language translation	covost-parquet/... only

Usage

This dataset is ideal for use in large-scale multilingual speech-to-text translation tasks, especially in distributed and high-performance computing environments. The parquet format enhances usability by minimizing I/O overhead, making it well-suited for high-throughput training.

import pyarrow.parquet as pq
from huggingface_hub import hf_hub_download

# Data (audio + texts):
p = hf_hub_download("meetween/mumospee_covost",
                    "covost-parquet/xx_en/covost2.de_en.validation.parquet",
                    repo_type="dataset")
table = pq.read_table(p)
# columns: audio, duration, transcription, translation

# Metadata only (same 13,511 rows, ~770× smaller):
m = hf_hub_download("meetween/mumospee_covost",
                    "metadata/covost2.de_en.validation.parquet",
                    repo_type="dataset")
manifest = pq.read_table(m).to_pandas()

Attribution

This dataset is based on the original CoVoST dataset, with modifications for I/O optimization by converting to parquet format. Please cite the original CoVoST dataset in any publications or projects using this dataset.

Changelog

2026-06-19

• Completed CoVoST 2 coverage — added the 9 previously missing language pairs: en_id (en→X), and ja_en, nl_en, pt_en, ru_en, sv-SE_en, ta_en, tr_en, zh-CN_en (X→en). The dataset now contains all 36 official CoVoST 2 pairs (15 en→X + 21 X→en), up from 27.
• Uniform schema with duration — every covost-parquet/*.parquet now includes a duration: double column. Existing pairs were re-uploaded from the same source preprocessing as the new ones to keep the schema consistent across the repo.
• metadata/ folder switched from CSV to Parquet — the per-pair / per-split metadata files (108 total) were regenerated as parquet, with duration: string to match the schema of meetween/mumospee. The legacy metadata/*.csv files were removed.
• README updated to document the 36-pair coverage, the new schema, and the role of each folder.

2024-11-25

• Initial release: 27 language pairs (14 en→X + 13 X→en) in parquet format, with per-pair metadata CSVs.