Update README.md

README.md

@@ -1,110 +1,304 @@
 ---
+license: other
+license_name: mixed-see-license-section
+task_categories:
+  - audio-to-audio
+language:
+  - en
+  - ko
+tags:
+  - speech-separation
+  - source-separation
+  - cocktail-party
+  - bilingual
+  - english
+  - korean
+  - speech
+pretty_name: MuSSeD Bilingual Speech Separation Mixtures
+size_categories:
+  - 10K<n<100K
 dataset_info:
-- config_name: 2spk
-  features:
-  - name: audio
-    dtype: audio
-  - name: file_id
-    dtype: string
-  - name: speaker_type
-    dtype: string
-  - name: meta
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 2161230124.0
-    num_examples: 10000
-  - name: dev
-    num_bytes: 281341657.3
-    num_examples: 1300
-  - name: test
-    num_bytes: 288195834.9
-    num_examples: 1300
-  download_size: 2704697738
-  dataset_size: 2730767616.2000003
-- config_name: 3spk
-  features:
-  - name: audio
-    dtype: audio
-  - name: file_id
-    dtype: string
-  - name: speaker_type
-    dtype: string
-  - name: meta
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 2608317239.0
-    num_examples: 10000
-  - name: dev
-    num_bytes: 337024379.8
-    num_examples: 1300
-  - name: test
-    num_bytes: 331952820.8
-    num_examples: 1300
-  download_size: 3196579806
-  dataset_size: 3277294439.6000004
-- config_name: 4spk
-  features:
-  - name: audio
-    dtype: audio
-  - name: file_id
-    dtype: string
-  - name: speaker_type
-    dtype: string
-  - name: meta
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 2958162977.0
-    num_examples: 10000
-  - name: dev
-    num_bytes: 378898175.9
-    num_examples: 1300
-  - name: test
-    num_bytes: 375917477.9
-    num_examples: 1300
-  download_size: 3610134728
-  dataset_size: 3712978630.8
-- config_name: scp
-  features:
-  - name: scp_filename
-    dtype: string
-  - name: content
-    dtype: string
-  splits:
-  - name: scp
-    num_bytes: 19744087
-    num_examples: 36
-  download_size: 3154315
-  dataset_size: 19744087
+  - config_name: 2spk
+    features:
+      - name: audio
+        dtype: audio
+      - name: file_id
+        dtype: string
+      - name: speaker_type
+        dtype: string
+      - name: meta
+        dtype: string
+    splits:
+      - name: train
+        num_bytes: 2161230124
+        num_examples: 10000
+      - name: dev
+        num_bytes: 281341657.3
+        num_examples: 1300
+      - name: test
+        num_bytes: 288195834.9
+        num_examples: 1300
+    download_size: 2704697738
+    dataset_size: 2730767616.2000003
+  - config_name: 3spk
+    features:
+      - name: audio
+        dtype: audio
+      - name: file_id
+        dtype: string
+      - name: speaker_type
+        dtype: string
+      - name: meta
+        dtype: string
+    splits:
+      - name: train
+        num_bytes: 2608317239
+        num_examples: 10000
+      - name: dev
+        num_bytes: 337024379.8
+        num_examples: 1300
+      - name: test
+        num_bytes: 331952820.8
+        num_examples: 1300
+    download_size: 3196579806
+    dataset_size: 3277294439.6000004
+  - config_name: 4spk
+    features:
+      - name: audio
+        dtype: audio
+      - name: file_id
+        dtype: string
+      - name: speaker_type
+        dtype: string
+      - name: meta
+        dtype: string
+    splits:
+      - name: train
+        num_bytes: 2958162977
+        num_examples: 10000
+      - name: dev
+        num_bytes: 378898175.9
+        num_examples: 1300
+      - name: test
+        num_bytes: 375917477.9
+        num_examples: 1300
+    download_size: 3610134728
+    dataset_size: 3712978630.8
+  - config_name: scp
+    features:
+      - name: scp_filename
+        dtype: string
+      - name: content
+        dtype: string
+    splits:
+      - name: scp
+        num_bytes: 19744087
+        num_examples: 36
+    download_size: 3154315
+    dataset_size: 19744087
 configs:
-- config_name: 2spk
-  data_files:
-  - split: train
-    path: 2spk/train-*
-  - split: dev
-    path: 2spk/dev-*
-  - split: test
-    path: 2spk/test-*
-- config_name: 3spk
-  data_files:
-  - split: train
-    path: 3spk/train-*
-  - split: dev
-    path: 3spk/dev-*
-  - split: test
-    path: 3spk/test-*
-- config_name: 4spk
-  data_files:
-  - split: train
-    path: 4spk/train-*
-  - split: dev
-    path: 4spk/dev-*
-  - split: test
-    path: 4spk/test-*
-- config_name: scp
-  data_files:
-  - split: scp
-    path: scp/scp-*
+  - config_name: 2spk
+    data_files:
+      - split: train
+        path: 2spk/train-*
+      - split: dev
+        path: 2spk/dev-*
+      - split: test
+        path: 2spk/test-*
+  - config_name: 3spk
+    data_files:
+      - split: train
+        path: 3spk/train-*
+      - split: dev
+        path: 3spk/dev-*
+      - split: test
+        path: 3spk/test-*
+  - config_name: 4spk
+    data_files:
+      - split: train
+        path: 4spk/train-*
+      - split: dev
+        path: 4spk/dev-*
+      - split: test
+        path: 4spk/test-*
+  - config_name: scp
+    data_files:
+      - split: scp
+        path: scp/scp-*
 ---
+
+# MuSSeD: Bilingual (English–Korean) Speech-Separation Mixtures
+
+A bilingual **speech-separation** corpus of synthetic multi-speaker mixtures built from
+English ([LibriTTS-R](https://www.openslr.org/141/)) and Korean read-speech sources.
+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 (equal share per split)
+- **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 unique (322 EN + 192 KO), **speaker-disjoint** across splits
+
+---
+
+## Reliability & Verification
+
+This dataset is designed for **trustworthy evaluation**. The following properties are
+programmatically verified by `dataset_audio.py` (raw log: `dataset_audio.0905.log`).
+
+### 1. Speaker-disjoint splits (no leakage)
+
+Speaker identity is parsed directly from the mixture filename (the token after each
+`EN_`/`KO_` tag) and cross-checked between splits. **No speaker appears in more than
+one split** — off-diagonal overlap is exactly 0, so a model can never be evaluated on a
+voice it saw during training.
+
+| Split | Unique speakers |
+|-------|----------------:|
+| train | 326 |
+| dev   | 100 |
+| test  | 88  |
+| **total** | **514** (322 EN + 192 KO) |
+
+Pairwise shared speakers: `train∩dev = 0`, `train∩test = 0`, `dev∩test = 0`.
+
+### 2. Representative splits (dev/test track train)
+
+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.
+
+![Split consistency](./figure/fig3_split_consistency.png)
+
+### 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.
+
+![Overlap distribution](./figure/fig4_overlap_distribution.png)
+
+### 4. Complete, machine-readable metadata
+
+Every mixture has a paired JSON with per-speaker ground truth (see schema below),
+enabling exact, reproducible analysis of language / gender / overlap / timing.
+
+---
+
+## Composition
+
+The share of bilingual (EN+KO) mixtures grows with speaker count, while EN-only /
+KO-only mixtures remain represented at every count:
+
+![Language composition](./figure/fig5_language_composition.png)
+
+---
+
+## Dataset Structure
+
+```
+Librittsr-KOR-n/
+├── train/ {2spk,3spk,4spk}/
+│   ├── raw/   <mixture_id>.wav      # 16 kHz mono mixture
+│   └── json/  <mixture_id>.json     # per-mixture ground-truth metadata
+├── dev/   {2spk,3spk,4spk}/ {raw,json}/
+├── test/  {2spk,3spk,4spk}/ {raw,json}/
+└── scp/   {2spk,3spk,4spk}/         # Kaldi-style file lists
+```
+
+**Filename convention.** A mixture filename is the concatenation of its source
+utterance ids, each of the form `EN_<spk>_...` or `KO_<spk>_...`. Example (4spk):
+
+```
+KO_5801_G1A2E7_HJH_001971_KO_5487_G1A4E7_KSH_004401_KO_5495_..._KO_5396_...wav
+                └ 4 sources → speakers KO_5801, KO_5487, KO_5495, KO_5396
+```
+
+### JSON schema (per mixture)
+
+```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"
+    }
+  ]
+}
+```
+
+---
+
+## Per-split statistics
+
+| Split / spk | # mix | Hours | Avg len | EN+KO | No-overlap | Mean overlap (>0) |
+|-------------|------:|------:|--------:|------:|-----------:|------------------:|
+| train/2spk | 10000 | 18.60 | 6.69 s | 36.1% | 21.4% | 0.589 |
+| train/3spk | 10000 | 21.98 | 7.91 s | 51.2% | 18.0% | 0.621 |
+| train/4spk | 10000 | 24.78 | 8.92 s | 59.4% | 15.3% | 0.651 |
+| dev/2spk   |  1300 |  2.40 | 6.65 s | 33.0% | 23.8% | 0.572 |
+| dev/3spk   |  1300 |  2.87 | 7.94 s | 51.1% | 18.3% | 0.627 |
+| dev/4spk   |  1300 |  3.26 | 9.02 s | 59.5% | 15.9% | 0.646 |
+| test/2spk  |  1300 |  2.48 | 6.87 s | 35.0% | 22.2% | 0.587 |
+| test/3spk  |  1300 |  2.87 | 7.95 s | 51.0% | 18.2% | 0.636 |
+| test/4spk  |  1300 |  3.25 | 9.00 s | 55.0% | 15.7% | 0.656 |
+
+---
+
+## Usage
+
+```python
+import json, glob, os, soundfile as sf
+
+root = "path/to/Librittsr-KOR-n"
+split, spk = "test", "2spk"
+
+for wav in sorted(glob.glob(f"{root}/{split}/{spk}/raw/*.wav")):
+    audio, sr = sf.read(wav)                     # 16 kHz mono mixture
+    meta = json.load(open(wav.replace("/raw/", "/json/").replace(".wav", ".json")))
+    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 with **different licenses**:
+
+- **English audio** originates from **LibriTTS-R** (CC BY 4.0).
+- **Korean audio** originates from a Korean read-speech corpus — **confirm and state
+  its license before public release.**
+
+Because of these mixed terms the top-level tag is set to `other`. Please verify the
+Korean-source license and update this section (and the `license`/`license_name` fields)
+accordingly prior to distribution.
+
+## Citation
+
+If you use this dataset, please cite LibriTTS-R and the Korean source corpus, and this
+release. *(Add BibTeX once the license/attribution is finalized.)*
+```bibtex
+@misc{bilingmousse2026,
+  title  = {MOUSSE: Multilingual Overlapped Speech Separation with Predictive Uncertainty-Guided Sampling},
+  year   = {2026},
+  note   = {2/3/4-speaker mixtures, speaker-disjoint splits}
+}
+```