File size: 4,831 Bytes
1621431
8b2c806
72127e7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1621431
8b2c806
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
---
license: other
dataset_info:
  features:
  - name: segment_id
    dtype: string
  - name: audio
    dtype:
      audio:
        decode: false
  - name: duration_seconds
    dtype: int64
  - name: segment_text
    dtype: string
  - name: cs_terms_list
    dtype: string
  - name: cs_terms_count
    dtype: int64
  - name: topic
    dtype: string
  - name: original_video_link
    dtype: string
  - name: original_video_title
    dtype: string
  - name: start_time
    dtype: string
  - name: end_time
    dtype: string
  splits:
  - name: train
    num_bytes: 14582022075
    num_examples: 11832
  - name: validation
    num_bytes: 2139515036
    num_examples: 1714
  - name: test
    num_bytes: 2026901460
    num_examples: 1614
  - name: hard
    num_bytes: 814798996
    num_examples: 658
  download_size: 18312886260
  dataset_size: 19563237567
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
  - split: hard
    path: data/hard-*
---
# 🩺 ViMedCSS: Vietnamese Medical Code-Switching Speech Dataset

## 📖 Overview
ViMedCSS is a Vietnamese medical speech dataset for code-switching ASR, where each utterance contains at least one non-Vietnamese (mainly English) medical term embedded in Vietnamese speech.

This dataset card is prepared from:
- The metadata files in `ViMedCSS-Metadata` (`train_set.csv`, `valid_set.csv`, `test_set.csv`, `hard_set.csv`)
- The paper: `Improving_Code_Switching_Detection_of_ASR_Models_for_Medical_Vietnamese_Speech (1).pdf`

From the paper, the full ViMedCSS corpus is reported as:
- 34.57 hours
- 16,576 utterances
- 5 medical topics
- 4 splits including a dedicated `hard` split for rare/unseen terms

From the provided CSV metadata in this release, the indexed subset contains:
- 15,818 rows
- 15,814 unique `segment_id`
- 32.64 hours total duration
- 16,581 total code-switched term occurrences
- 889 distinct code-switched medical terms

## 📊 Dataset Statistics
### Split Statistics (from `ViMedCSS-Metadata`)
| Split | # Rows | Duration (hours) | Avg duration (s) | Total CS terms |
|---|---:|---:|---:|---:|
| train | 11,832 | 24.30 | 7.39 | 12,314 |
| validation | 1,714 | 3.57 | 7.49 | 1,814 |
| test | 1,614 | 3.39 | 7.56 | 1,695 |
| hard | 658 | 1.38 | 7.57 | 758 |
| **Total** | **15,818** | **32.64** | **7.43** | **16,581** |

### Topic Statistics (from `ViMedCSS-Metadata`)
| Topic | # Rows | Duration (hours) | Total CS terms |
|---|---:|---:|---:|
| Medical Sciences | 6,836 | 14.68 | 7,459 |
| Pathology & Pathogens | 4,827 | 10.00 | 4,951 |
| Treatments | 1,969 | 3.80 | 1,985 |
| Nutrition | 1,155 | 2.14 | 1,155 |
| Diagnostics | 1,031 | 2.02 | 1,031 |

## 🧾 Data Fields
Each row in metadata corresponds to one segment audio file, where:
- `segment_id` maps to `segment_id.wav` (for example: `Med_CS-100-17` -> `Med_CS-100-17.wav`)

Main fields:
- `segment_id`: utterance identifier
- `duration_seconds`: utterance duration
- `segment_text`: Vietnamese transcript containing code-switched term(s)
- `cs_terms_list`: semicolon-separated code-switched terms
- `cs_terms_count`: number of code-switched terms in the utterance
- `topic` (or `Topic` in one CSV): medical topic label
- `original_video_link`: source video URL
- `original_video_title`: source video title
- `start_time`, `end_time`: segment boundaries in source audio/video

When loaded from Hugging Face, an `audio` column is available with waveform bytes/path in the standard 🤗 Datasets `Audio` format.

## 🔽 How to Load
Load directly with 🤗 Datasets:

```python
from datasets import load_dataset

dataset = load_dataset("tensorxt/ViMedCSS")
print(dataset)
```

Clone with Git LFS:

```bash
git lfs install
git clone https://huggingface.co/datasets/tensorxt/ViMedCSS
```

## 📝 Notes
- The paper reports the full corpus statistics (34.57h, 16,576 utterances).
- The CSV metadata bundled here describes a smaller processed subset (32.64h, 15,818 rows).
- The `hard` split is intended for evaluating rare/unseen code-switched medical terms, following the paper’s benchmark setup.

## 📜 License
The paper states that data are collected from publicly available YouTube content for research purposes, and the medical dictionary resource used in construction is under institutional intellectual property licensing.

Please verify usage rights for your setting before redistribution or commercial use.

## 🙏 Citation
If you use ViMedCSS, please cite:

```bibtex
@misc{nguyen_vimedcss,
  title={ViMedCSS: A Vietnamese Medical Code-Switching Speech Dataset \& Benchmark},
  author={Nguyen, Tung X. and Vo, Nhu and Nguyen, Giang-Son and Hoang, Duy Mai and Huynh, Chien Dinh and Jauregi Unanue, Inigo and Piccardi, Massimo and Buntine, Wray and Le, Dung D.},
}
```