Gong1212/CoreaSpeech-plus

CoreaSpeech+

CoreaSpeech+ is a unified Korean speech dataset repository for text-to-speech research, speech synthesis, validation/test split management, and speech corpus distribution.

This repository organizes the CoreaSpeech dataset family, including the 700-hour CoreaSpeech corpus, additional training data for the 2000-hour configuration, validation data, and the Korean Universal Testset.

Overview

CoreaSpeech+ is designed to support reproducible Korean speech generation research.

The dataset family consists of:

Subset	Description	Status
CoreaSpeech 700h	Base Korean speech corpus selected through the CoreaSpeech pipeline	Available
CoreaSpeech 2000h	Logical combined corpus consisting of CoreaSpeech 700h and additional training data	Available
Validation	Validation split for Korean TTS training and model selection	Available
Korean Universal Testset	Evaluation benchmark with clean, noisy, and numeric subsets	Available

Dataset Statistics

Subset	Audio files	Metadata rows	Text rows	Approx. hours	Notes
700h	168,790	168,790	168,000	697.65	Base training subset
2000h	697,168	697,168	696,378	2,154.03	Combined set: 700h base + additional training data
Validation	100	100	100	-	Validation split
Test	600 wav files	300 rows	300 paired rows	-	clean/noisy/numeric, 100 rows each

Dataset Composition

CoreaSpeech 700h

The 700-hour subset corresponds to the base CoreaSpeech corpus.

According to the related paper, CoreaSpeech is a Korean speech corpus constructed through Korean-specific preprocessing and Jamo-based coreset selection. It contains approximately 700 hours of speech, 168,790 samples, and 21,449 speakers.

CoreaSpeech 2000h

The 2000-hour configuration is a logical combined corpus.

It consists of:

CoreaSpeech 700h
+ additional training data
= CoreaSpeech 2000h

To avoid duplicate storage, the 2000h manifest should reference audio files from both:

data/train_700h/
data/train_extra_1300h/

rather than storing a separate duplicated data/train_2000h/ directory.

Validation

The validation split contains:

Item	Count
Audio files	100
Metadata rows	100
Text rows	100

Korean Universal Testset

The Korean Universal Testset is organized into three evaluation subsets:

Test subset	Metadata rows	Audio files	Description
clean	100	200	Clean speech evaluation subset
noisy	100	200	Robustness evaluation subset under noisy acoustic conditions
numeric	100	200	Numeric and special-expression evaluation subset
total	300	600	Korean Universal Testset

The testset contains 300 metadata rows and 600 wav files in total.

The clean subset evaluates synthesis under relatively clean acoustic conditions.
The noisy subset evaluates robustness under challenging acoustic conditions.
The numeric subset evaluates Korean numeric pronunciation, including numerals, floating-point numbers, special symbols, abbreviations, and English-unit expressions.

Audio Specifications

Field	Value
Audio format	WAV
Experimental sampling rate	24,000 Hz
Released sampling rate	To be verified
Channels	To be verified
Bit depth	To be verified
Language	Korean
Speaker count	21,449 speakers for the 700h CoreaSpeech subset

Note: The related paper reports that training samples were downsampled to 24 kHz for experiments. The final released audio specification should be verified from the uploaded files.

Planned Repository Structure

CoreaSpeech-plus/
├── README.md
├── VERSION
├── manifests/
│   ├── 700h.tsv
│   ├── 2000h.tsv
│   ├── validation.tsv
│   └── test.lst
├── data/
│   ├── train_700h/
│   ├── train_extra_1300h/
│   ├── validation/
│   └── test/
└── docs/
    ├── access_policy.md
    ├── dataset_structure.md
    ├── metadata_schema.md
    └── changelog.md

Manifest Design

The manifest files define how each subset should be loaded.

manifests/700h.tsv

contains metadata for the 700h base subset.

manifests/extra_1300h.tsv

contains metadata for the additional training data used to construct the 2000h configuration.

manifests/2000h.tsv

contains the combined metadata for the full 2000h configuration. This manifest references files from both data/train_700h/ and data/train_extra_1300h/.

manifests/validation.tsv

contains metadata for the validation split.

manifests/test.lst

contains paired metadata for the Korean Universal Testset.

WebDataset-style Archive Format

Large-scale audio files are distributed as TAR shards.

Example structure:

data/train_700h/
├── train_700h-000000.tar
├── train_700h-000001.tar
└── ...

data/train_extra_1300h/
├── train_extra_1300h-000000.tar
├── train_extra_1300h-000001.tar
└── ...

data/validation/
├── validation-000000.tar
└── ...

data/test/
├── test_clean-000000.tar
├── test_noisy-000000.tar
└── test_numeric-000000.tar

Each TAR shard should contain audio files and corresponding metadata entries with aligned sample prefixes:

train_700h-000000.tar
├── 000000000.wav
├── 000000000.json
├── 000000001.wav
├── 000000001.json
└── ...

For the testset, each metadata row may correspond to paired audio files depending on the evaluation protocol.

Example:

test_clean-000000.tar
├── clean_000000_prompt.wav
├── clean_000000_target.wav
├── clean_000000.json
├── clean_000001_prompt.wav
├── clean_000001_target.wav
├── clean_000001.json
└── ...

Metadata Schema

CoreaSpeech+ provides metadata in tabular formats rather than JSON schema files.

Training and validation metadata are provided as TSV files, while the Korean Universal Testset follows a paired-list format.

Training Metadata: 700h

The 700h metadata is provided as a TSV file.

Recommended file:

manifests/700h.tsv

Columns:

Column	Type	Description
relpath	string	Relative path to the audio file
duration	float	Audio duration in seconds
text	string	Transcription text
source	string	Optional source or partition identifier

Recommended format:

relpath	duration	text	source
data/train_700h/train_700h-000000.tar/000000000.wav	3.42	한국어 전사 문장	source_name

If source contains internal or sensitive provenance information, it may be omitted from the public manifest.

Training Metadata: 2000h

The 2000h metadata is provided as a TSV file.

Recommended file:

manifests/2000h.tsv

The 2000h split is a logical combined corpus consisting of the 700h base subset and additional training data. Therefore, the metadata should include dataset_part to indicate which subset each sample belongs to.

Columns:

Column	Type	Description
relpath	string	Relative path to the audio file
duration	float	Audio duration in seconds
text	string	Transcription text
dataset_part	string	Subset identifier, such as coreaspeech_700h or coreaspeech_extra_1300h
source	string	Optional source or partition identifier

Recommended format:

relpath	duration	text	dataset_part	source
data/train_700h/train_700h-000000.tar/000000000.wav	3.42	한국어 전사 문장	coreaspeech_700h	source_name
data/train_extra_1300h/train_extra_1300h-000000.tar/000000000.wav	4.18	추가 학습 데이터 문장	coreaspeech_extra_1300h	source_name

The 2000h.tsv manifest should reference files from both:

data/train_700h/
data/train_extra_1300h/

This avoids duplicate storage while preserving the logical 2000h training configuration.

Validation Metadata

The validation metadata is provided as a TSV file.

Recommended file:

manifests/validation.tsv

Columns:

Column	Type	Description
relpath	string	Relative path to the validation audio file
duration	float	Audio duration in seconds
text	string	Transcription text

Recommended format:

relpath	duration	text
data/validation/validation-000000.tar/000000000.wav	3.57	검증용 한국어 문장

Korean Universal Testset Metadata

The Korean Universal Testset follows a paired metadata format.

Recommended file:

manifests/test.lst

Format:

prompt_wav|prompt_duration|prompt_text|target_wav|target_duration|target_text

Columns:

Column	Type	Description
prompt_wav	string	Relative path to the prompt/reference audio
prompt_duration	float	Prompt audio duration in seconds
prompt_text	string	Text corresponding to the prompt audio
target_wav	string	Relative path to the target audio
target_duration	float	Target audio duration in seconds
target_text	string	Target text for evaluation

Recommended format:

data/test/clean/clean_000000_prompt.wav|3.12|프롬프트 문장|data/test/clean/clean_000000_target.wav|4.21|합성 대상 문장

The testset consists of three subsets:

Subset	Rows	WAV files	Description
clean	100	200	Clean speech evaluation subset
noisy	100	200	Noisy speech evaluation subset
numeric	100	200	Numeric and special-expression evaluation subset
total	300	600	Korean Universal Testset

If the test metadata is stored as a single file, the subset can be inferred from the file path, such as data/test/clean/, data/test/noisy/, and data/test/numeric/.

Access Policy

CoreaSpeech+ is distributed under controlled access due to licensing, copyright, and research-use restrictions.

Users may be required to provide:

• Name
• Affiliation
• Email
• Intended use
• Agreement to the terms of use

Access may be approved, rejected, or revoked by the maintainer.

Intended Use

CoreaSpeech+ is intended for academic and non-commercial research, including:

• Korean text-to-speech model training
• Korean speech synthesis research
• Multi-speaker TTS research
• Low-resource or cross-lingual speech generation experiments
• Speech corpus analysis and preprocessing
• Validation and evaluation of Korean TTS systems

Prohibited Use

The dataset must not be used for:

• Redistribution of the original audio files
• Commercial services without explicit permission
• Speaker impersonation or identity misuse
• Fraud, misinformation, or harmful applications
• Any use that violates copyright, privacy, or institutional policies

Download

After access is approved, users can download the repository with the Hugging Face CLI:

huggingface-cli login

huggingface-cli download Gong1212/CoreaSpeech-plus \
  --repo-type dataset \
  --local-dir ./CoreaSpeech-plus

Or with Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Gong1212/CoreaSpeech-plus",
    repo_type="dataset",
    local_dir="./CoreaSpeech-plus"
)

Related Paper

CoreaSpeech+ is organized as a unified Hugging Face repository for managing and distributing Korean speech resources related to CoreaSpeech.

If you use this dataset, please cite the related paper:

@article{kwon2026coreaspeech,
  title={CoreaSpeech: Korean Speech Corpus via Jamo-based Coreset Selection for Efficient and Robust Korean Speech Generation},
  author={Kwon, Ki-Joong and So, Junho and Lee, Sang-Hoon},
  journal={Advances in Neural Information Processing Systems},
  volume={38},
  year={2026}
}