syshim/MuSSeD-EnKo

MuSSeD: Multilingual Speech Separation Dataset (English, Korean)

A bilingual speech-separation corpus of synthetic multi-speaker mixtures built from English (LibriTTS-R) and Korean read-speech sources (AI-Hub Multi-Speaker Speech Synthesis corpus). Each mixture concatenates/overlaps 2, 3, or 4 single-speaker utterances, and every mixture ships with rich per-speaker JSON metadata (speaker id, gender, language, timing, transcription, overlap ratio, source utterance id).

• Audio: 16 kHz, mono, 16-bit PCM WAV
• Mixtures: 37,800 total — 30,000 train / 3,900 dev / 3,900 test
• Speaker counts: 2spk, 3spk, 4spk
• Total duration: ≈ 82.5 h (train ≈ 65.4 h, dev ≈ 8.5 h, test ≈ 8.6 h)
• Languages: English (ENG) + Korean (KOR), including bilingual mixtures
• Speakers: 514 speakers

---

Reliability & Verification

MuSSeD is built for leakage-free, representative evaluation: splits are speaker-disjoint, dev/test mirror the train distribution, overlap spans the full difficulty range, and every mixture ships with complete per-speaker ground truth — each property is summarized below.

1. Speaker-disjoint splits

Speaker identity is parsed directly from the mixture filename (the token after each EN_/KO_ tag) and cross-checked between splits.

Split	Unique speakers
train	326
dev	100
test	88
total	514 (322 EN + 192 KO)

2. Representative splits

The dev and test splits follow the same distributions as train for the properties that matter for separation difficulty — bilingual share and overlap share are nearly identical across splits, so validation/test numbers are representative of the training distribution.

3. Controlled, full-range overlap

Overlap is not concentrated at one difficulty. Mixtures span the full range from clean (no overlap) to fully overlapped, with the harder end growing as speaker count increases (2-4 spk), which is the expected and desired behaviour.

4. Metadata

Every mixture has a paired JSON with per-speaker ground truth, enabling exact, reproducible analysis of language / gender / overlap / timing.

---

Composition

The share of bilingual mixtures grows with speaker count, while English-only / Korean-only mixtures remain represented at every count:

Across the 514 unique speakers, the male / female balance is:

---

Dataset Structure

MuSSeD-EnKo/
├── train/ {2spk,3spk,4spk}/
│   ├── raw/   <mixture_id>.wav
│   └── json/  <mixture_id>.json
├── dev/   {2spk,3spk,4spk}/ {raw,json}/
└── test/  {2spk,3spk,4spk}/ {raw,json}/

Filename convention. A mixture filename is the concatenation of its source utterance ids, each of the form EN_<spk>_... or KO_<spk>_... — e.g. a 4spk file KO_5801_..._KO_5487_..._KO_5495_..._KO_5396_....wav encodes speakers KO_5801, KO_5487, KO_5495, KO_5396.

JSON

{
  "mixture_id": "EN_260_123286_000027_000001_EN_4992_41806_000003_000002",
  "speaker_count": 2,
  "overlapped_ratio": 0.0,
  "total_length_sec": 14.44,
  "speakers": [
    {
      "speaker_id": "260",
      "speaker_name": "Brad Bush",
      "gender": "Male",
      "language": "ENG",
      "length_sec": 2.48,
      "start_time_sec": 0.0,
      "end_time_sec": 2.48,
      "transcription": "I had rather not disturb him while he is quiet.",
      "original_file_id": "/test-clean/260/123286/260_123286_000027_000001"
    }
  ]
}

---

Usage

from datasets import load_dataset
import json

# config = "2spk" | "3spk" | "4spk"
ds = load_dataset("syshim/MuSSeD-EnKo", "2spk", split="test")

ex      = ds[0]
audio   = ex["audio"]              # {"array": np.ndarray, "sampling_rate": 16000}
meta    = json.loads(ex["meta"])   # per-mixture ground truth
n_src   = meta["speaker_count"]
overlap = meta["overlapped_ratio"]
langs   = [s["language"] for s in meta["speakers"]]
# ... train / evaluate your separation model ...

---

License

This dataset is derived from multiple sources:

• English audio originates from LibriTTS-R (CC BY 4.0).
• Korean audio originates from a Multi-speaker Speech Synthesis Data by National Information Society Agency (NIA)

Because of these mixed terms the top-level tag is set to other.

Citation

If you use this dataset, please cite LibriTTS-R, the Korean source corpus, and this release:

@misc{mousse2026,
  title  = {MOUSSE: MULTILINGUALOVERLAPPEDSPEECHSEPARATIONWITHPREDICTIVE UNCERTAINTYESTIMATION},
  year   = {2026}
}