This corpus is derived from several Korean speech datasets (AI-Hub, KlecSpeech, KsponSpeech, CoreaSpeech, Emilia, kosp2e) that each retain their own license. The upstream sources DO NOT permit redistribution — access here is granted on manual review for research use only, and you must hold your own agreements with the upstream providers where their terms require it.

S2ST-ko-hibiki — Korean→English Simultaneous Speech-to-Speech Translation

A large-scale corpus for Korean → English simultaneous (streaming) speech-to-speech translation, built to fine-tune a Hibiki-style streaming model. Each item pairs real Korean source speech with an English translation (text) and synthesized English speech, plus Mimi neural-codec tokens for both languages and word-level alignment timestamps — so the whole set loads directly into a streaming S2S training loop.

Backs the paper TransOrder: Reinforcement Learning for Word Reordering in Simultaneous Speech-to-Speech Translation (SAIL, Ajou University).

Access is gated (manual approval). The upstream corpora state redistribution is not permitted; see Licensing.

Composition

Training-ready segments in packed_v2 (post-QC, evaluation ids excluded). Hours are of the synthesized English audio unless noted.

Source key	Official dataset	AI-Hub ID	Content	Speakers	Segments	EN hours
`aihub130`	한국인 대화음성 (Korean conversational speech)	130	2-person dialogue (economy/daily/commerce)	(source meta)	873,839	3,009
`klec`	KlecSpeech 한국어 강의음성 (Korean lecture speech)	KlecSpeech	20-domain lectures, 14 difficulty levels	7,942	973,603	4,777
`aihub71481`	전문분야 심층인터뷰 (expert in-depth interviews)	71481	host/panel interviews (rich demographic meta)	~2,679 interviews ⚠️	291,773	1,511
`kspon`	KsponSpeech 한국어 음성	KsponSpeech	spontaneous free speech	2,000 (official)	204,969	739
`aihub637`	감성·발화스타일별 음성합성 (SKT)	637	9 emotions (neutral/anger/sadness/joy…)	508	67,989	256
`coreaspeech`	CoreaSpeech	(HuggingFace)	Korean speech	799	116,603	600
`aihub71668`	라이브 스트리밍 영어 통번역 (live-stream EN interpretation)	71668	YouTuber native EN + KO clone-voice; KO+EN both real	105	197,686	865
`aihub031`	방송콘텐츠 한-영 통·번역 음성 (broadcast interpretation)	71379	documentary/entertainment; KO + native EN	59	114,778	395
`aihub71261`	한-영·한-중 음성발화 (KO-EN/KO-ZH utterances)	71261	native EN speech utterances	(source meta)	26,842	94
Total					2,868,082	12,245.6

Totals across packed_v2: 2,868,082 segments · 6,275,250 members · KO 10,418.3 h · EN 12,245.6 h (notes/datacard_stats.json).

Dataset structure

S2ST-ko-hibiki/
├── README.md
├── ko/                    # Korean source audio (tar.zst shards) — one folder per source key above
├── en/                    # English synthesized audio (OmniVoice)
├── packed_v2/             # ★ training-ready pack: KO+EN Mimi tokens + EN/KO text + delay anchors + ids
├── align/
│   ├── ko/                # KO word-level forced-alignment timestamps (MMS)
│   └── en/                # EN word-level forced-alignment timestamps (MMS)
├── code/                  # data-pipeline scripts (pack / synth / qc / align / train / eval / rl)
├── splits/                # frozen dev/test ids + Emilia eval set
└── stats/                 # per-source counts, durations, QC dumps

packed_v2/ is the single canonical training artifact; per source it holds <key>.tok.bin (Mimi RVQ tokens, KO+EN interleaved per frame), <key>.txt.bin (English translation, SentencePiece), <key>.kotext.txt (Korean transcript), <key>.rows.npz (offsets + delay anchors), <key>.ids.txt, and <key>.meta.json.

`ko/` folder → source mapping

Korean source audio is stored as tar.zst shards under ko/<source key>/. Folder names follow the source keys above. Three keys differ from their earlier raw-backup folder names:

`ko/` folder	earlier raw-backup name	note
`klec`	`raw_aihub_ebs`	KlecSpeech (lecture speech) — not a broadcast set
`aihub031`	`raw_aihub71379_broadcast`	same corpus as AI-Hub 71379 (broadcast interpretation)
`aihub71668`	`raw_no_kor_data`	folder was mis-named; holds AI-Hub 71668 (+ an Emilia-KO latin subset)

ko/ also carries held-out-only Korean audio (kosp2e, emilia_ko) that is never used for training. The raw audio for aihub71261 is not included in this backup (only its tokens/text/alignment are in packed_v2).

How it was built

Source speech — real Korean recordings from the corpora above.
Translation (KO→EN text) — original English references where available, plus Hunyuan-MT-7B re-translation, scored with CometKiwi QE and merged keep-max (higher-QE of original vs. re-translation per segment).
English speech synthesis — rendered with OmniVoice (lang=en, speed=0.9, target duration matched to the source segment). The accented ("Konglish") delivery is an intentional, documented property, not a defect.
Neural-codec tokenization — both Korean and English audio encoded to Mimi RVQ tokens (12.5 Hz); English text tokenized with SentencePiece.
Forced alignment — Korean word-level timestamps via MMS forced alignment (used as the latency-reward clock in the paper).
Quality control — automatic language-ID screening drops segments whose synthesis came out non-English; text-contaminated segments (~0.97%) removed; foreign+high-WER cases re-synthesized once, then re-QC'd.
Packing — merged into the streaming grid (packed_v2) with per-segment delay anchors; evaluation ids held out.

Translation quality (measured)

QE (CometKiwi-DA) on the full ~7.4 M sentences: per-dataset mean ≈ 0; low-quality share (QE < −0.3) 0.00–1.26% (weakest: coreaspeech 1.26%, aihub71261 0.60%).
LaBSE cosine (KO↔EN, 2 k sample/dataset): mostly mean 0.75–0.78. Outliers with mis-alignment tails: aihub71261 (19.3% below 0.5), aihub637 (9.4%) — aihub71261 also shows the largest KO/EN duration mismatch (48%).

Splits

Train — packed_v2 (source-level split; evaluation material removed via splits/exclude_from_train.*, 80,755 segments).
Internal dev / test — frozen id lists in splits/ (seed 46, source-disjoint; two broadcast programs held fully unseen).
Emilia eval — splits/emilia_eval/: 500 short (3–15 s) + 500 long (25–90 s) reference-free clips (DNSMOS-filtered).
External benchmarks (FLEURS/FLoRes KO→EN, kosp2e) are referenced in the paper but not used for training.

Licensing

⚠️ This corpus aggregates material under several different upstream licenses, and the upstream providers DO NOT permit redistribution. It is published with manual-approval gating for research use only.

AI-Hub (aihub130, aihub637, aihub031/71379, aihub71481, aihub71668, aihub71261), KlecSpeech (klec), KsponSpeech (kspon): governed by their AI-Hub terms — https://www.aihub.or.kr. Approved recipients are expected to hold their own AI-Hub agreements.
CoreaSpeech (coreaspeech), Emilia (emilia_ko), kosp2e: governed by their original release terms.
Derived artifacts produced by us — English translations, OmniVoice English speech, Mimi tokens, and MMS alignments — are provided for research use. Mimi tokens are a lossy, non-invertible representation.

If you are a rights holder with concerns about any content, please open a discussion on this repository.

Intended use & limitations

Intended use: research on simultaneous / streaming speech-to-speech translation, KO→EN translation, and neural-codec speech modeling.
English speech is synthetic (OmniVoice), with intentional accented delivery — a translation target for training, not a natural-speech corpus.
Translations are largely machine-generated (except native-EN sources); expect residual translation noise (notably aihub71261).
Broadcast sources (aihub031, aihub71261) contain multi-speaker, possibly real-person audio — review privacy before any downstream release. Not for speaker identification or de-anonymization.

Citation

@inproceedings{transorder2026,
  title     = {TransOrder: Reinforcement Learning for Word Reordering in Simultaneous Speech-to-Speech Translation},
  author    = {⚠️ CONFIRM author list},
  booktitle = {⚠️ CONFIRM venue},
  year      = {2026}
}

Contact

SAIL, Ajou University.

Downloads last month: 52

Total file size:

5.07 TB

You need to agree to share your contact information to access this dataset