You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This repository shares reference examples with research associates and colleagues. It includes the original Emilia dataset, whose original Terms of Access follow. The Emilia-YODAS dataset is not included in this bank.
Terms of Access: The researcher has requested permission to use the Emilia dataset, the Emilia-Pipe preprocessing pipeline, and the Emilia-Yodas dataset. In exchange for such permission, the researcher hereby agrees to the following terms and conditions:
- The researcher shall use the Emilia dataset under the CC-BY-NC license and
the Emilia-YODAS dataset under the CC-BY license. - The authors make no representations or warranties regarding the datasets,
including but not limited to warranties of non-infringement or fitness for
a particular purpose. - The researcher accepts full responsibility for their use of the datasets and
shall defend and indemnify the authors of Emilia, Emilia-Pipe, and
Emilia-Yodas, including their employees, trustees, officers, and agents,
against any and all claims arising from the researcher's use of the datasets,
including but not limited to the researcher's use of any copies of copyrighted
content that they may create from the datasets. - The researcher may provide research associates and colleagues with access
to the datasets, provided that they first agree to be bound by these terms
and conditions. - The authors reserve the right to terminate the researcher's access to the
datasets at any time. - If the researcher is employed by a for-profit, commercial entity, the
researcher's employer shall also be bound by these terms and conditions,
and the researcher hereby represents that they are fully authorized to enter
into this agreement on behalf of such employer.
Log in or Sign Up to review the conditions and access this dataset content.
Authoritative production release: reviewed schema 2
Use this metadata patch with the unchanged audio from all three original ZIP parts. Production selection uses the reviewed flat manifest and immutable schema-2 catalog below. The older ZIP part manifests are historical packaging snapshots and no longer define production eligibility.
- Production reference clips after voice/audio deduplication: 775,580.
- Reviewed accepted manifest: 787,581 records (784,407 accepted, 3,174 corrected).
- Quarantined: 11,690; rejected: 603.
- Reviewed synthetic bilingual pairs retained: 771.
- Supplemental manual review quarantined 3 records; decisions:
exports/source_language_qc_20261019_v3/reviewed_exclusions.jsonl, SHA2569de4679dd15d88e8b558cc1302f0bb25d92f28df854617d0fcc458931898c23d.
| Production language | Reference clips |
|---|---|
| en-US | 407,760 |
| zh-CN | 367,820 |
The minimum reference duration is 5 seconds, with no arbitrary 12-second maximum. Source metadata, the source transcript and SHA-bound cached ASR are checked for one primary supported language. Small foreign terms are allowed; significant mixed, unknown and unsupported language is excluded. Cached forced-language ASR is not independent acoustic language identification or a listening review.
Catalog ID: 2e05c396b96352134561d1f4e4dd6cd28f754f40e4e96d4b1301290e5278a883. Language policy SHA256:
d6c6e9b381f03e76316bc27fce1840939e68b911b26505772483a226119a00f1. Reviewed manifest:
exports/source_language_qc_20261019_v3/accepted.jsonl. Exact counts, exclusions and byte hashes are
in PRODUCTION_STATUS.json and the patch's SHA256SUMS.
hf auth login
hf download EricGC/Voice_bank voice_bank_reviewed_schema2_source_language_qc_20261019_v3.tar.gz voice_bank_reviewed_schema2_source_language_qc_20261019_v3.tar.gz.sha256 --repo-type dataset --local-dir .
sha256sum -c voice_bank_reviewed_schema2_source_language_qc_20261019_v3.tar.gz.sha256
mkdir -p reviewed-stage
tar -xzf voice_bank_reviewed_schema2_source_language_qc_20261019_v3.tar.gz -C reviewed-stage
(cd reviewed-stage && sha256sum -c SHA256SUMS)
cat reviewed-stage/INSTALL_REVIEWED.md
Follow the installation commands in INSTALL_REVIEWED.md to back up and replace the
root catalog pair while preserving all bank-relative reviewed paths. Use the
matching schema-2 Breeze workers; allocate the complete ordered session manifest
before sharding. Random shuffling uses no replacement within eligible pools;
persistent compact cursor state continues balanced voice use across runs.
Sources have no equal-use quotas. Old schema-1 catalogs fail with a migration
error. The fixed production assistant Serena references remain unchanged.
Existing ZIP bytes, licenses and automatic research-access gating are unchanged. The original dataset description and source terms follow as historical release documentation; its older availability/manifest counts are not production totals.
Chinese and English Voice Reference Bank
One automatically screened reference utterance per source speaker/group ID, plus 1,000 Breeze-designed bilingual voice pairs (2,000 synthetic clips). Source IDs are not globally deduplicated human identities. Different voice descriptions do not guarantee different acoustic voices.
Three-part release
Parts 01 and 02 are immutable, balanced subsets of the accepted snapshot taken when publication started. The initial snapshot uses a deterministic hash of each source speaker ID or synthetic pair ID to assign three groups. A synthetic Chinese/English pair always stays together. Part 03 is produced only after all source collection, automatic QC and identity processing finish; it contains every final accepted voice/group absent from the first two parts, including the initially reserved third group and all later additions.
| Part | Reference clips | Publication stage |
|---|---|---|
| 01 | 245963 | uploaded |
| 02 | 245778 | uploaded |
| 03 | 308133 | uploaded |
Each ZIP is independently usable and uses standard ZIP64 with lossless deflate.
No audio is resynthesized or changed for publication. Extract all parts into
the same directory. Relative audio paths are preserved (public/... and
synthetic_breeze1000/...); per-part manifests are
manifests/part_01.jsonl, manifests/part_02.jsonl and manifests/part_03.jsonl.
Each part also has a standalone gzipped manifest, a JSON summary and a ZIP
SHA256 sidecar. Manifests include transcription, language, source repository
and pinned revision, source license, waveform SHA256, quality evidence and
available voice attributes. Synthetic pair provenance and its current
identity evidence are retained. Machine-specific absolute paths are removed.
Quality and identity evidence
Public references are single complete 6–13 second utterances, converted to 24 kHz mono in the collection pipeline. They pass DNSMOS signal >=3.5, background >=3.8, overall >=3.3 and independent Whisper CER/WER <=5%, together with waveform checks. Synthetic references are 8–12 seconds and require both languages to pass, a current Chinese-to-English cloning reference hash and an end-of-transcript screen. These are automatic checks, not human listening approval. Acoustic identity evidence is heuristic; neither source diarization IDs nor model thresholds establish globally unique people or voices.
Sources and licensing
| Content | Source | Terms |
|---|---|---|
| Original Emilia ZH/EN | amphion/Emilia-Dataset | CC BY-NC 4.0 and original Terms of Access |
| AISHELL-3 | OpenSLR 93 | Apache-2.0 |
| LibriTTS-R | OpenSLR 141 | CC BY 4.0 |
| GLOBE V2 | MushanW/GLOBE_V2 | CC0-1.0 |
| Breeze synthetic examples | BreezeBlue/Breeze-TTS-2 | BreezeBlue Research and Non-Commercial License v1.1 |
The placeholder Apache license originally used to create this repository does
not license all its audio. See LICENSES/ and each manifest's source/license.
Original Emilia references are provided for non-commercial research subject
to the access terms above. The Breeze examples retain their model-output
conditions, including restrictions on commercial use and training/improving
other generative models; consult the full Breeze license. The model's source
code Apache license does not replace its model/output license. Source credits,
original IDs and upstream rights remain with their respective contributors.
The original Emilia access terms are reproduced in
LICENSES/EMILIA_ORIGINAL_DATASET_CARD.md. Access is automatically granted
after a researcher supplies the original fields and expressly accepts those
same terms. This is a reference subset of original Emilia, not Emilia-YODAS.
Download and verify
The archives can exceed 50 GB. Use the current Hugging Face client with Xet support instead of older plain-HTTP download clients:
pip install -U huggingface_hub hf_xet
hf auth login
hf download EricGC/Voice_bank --repo-type dataset --local-dir ./Voice_bank
cd Voice_bank
sha256sum -c voice_bank_part_01.zip.sha256
unzip voice_bank_part_01.zip
Repeat verification/extraction for the other available parts. Part 03 appears
only when final construction and validation have finished. UPLOAD_STATUS.json
records the currently published parts; it does not count unfinished archives
as uploaded. Large files are verified against the server's LFS SHA256 after
their upload commits complete.
- Downloads last month
- 417