---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: caption
    dtype: string
  - name: start
    dtype: float64
  - name: end
    dtype: float64
  - name: sensevoice_caption
    dtype: string
  - name: sensevoice_words
    list:
    - name: word
      dtype: string
    - name: start
      dtype: float64
    - name: duration
      dtype: float64
  - name: audio
    dtype:
      audio:
        sampling_rate: 16000
  - name: uploader_id
    dtype: string
  - name: video_id
    dtype: string
  splits:
  - name: train
    num_bytes: 390072658.15165097
    num_examples: 56484
  - name: test
    num_bytes: 6968997.504349036
    num_examples: 1000
  download_size: 385165068
  dataset_size: 397041655.656
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
---



# YouTube ASR Caption Dataset (Cantonese)

This dataset was built from YouTube videos with **manually provided captions** in Cantonese. We used [SenseVoice](https://github.com/sensevoice/sensevoice) to re-transcribe the audio and filtered segments to build a high-quality collection of **audio-caption pairs**.

### What’s included
- Segments where the ASR output is **identical** to the original caption — likely clean.
- Segments where differences are only **homophones (同音字)** or **English words** — likely ASR mistakes.

This combination supports both:
- **ASR error analysis and correction**
- **Training clean speech-to-text models**

### Dataset Info
- **Total duration**: ~35 hours
- **Sampling rate**: 16 kHz
- **Audio format**: `.mp3`
- **Unfiltered data**: available under `creator/{video_id}/*.mp3`

### Features

| Name                | Type                                   |
|---------------------|----------------------------------------|
| `id`               | `string`                               |
| `caption`          | `string`                               |
| `start`, `end`     | `float64` (seconds in original audio)  |
| `sensevoice_caption` | `string`                             |
| `sensevoice_words` | list of { `word`, `start`, `duration` } |
| `audio`            | `Audio(sampling_rate=16000)`           |
| `uploader_id`      | `string`                               |
| `video_id`         | `string`                               |

### Splits

| Split | # Examples | Size    |
|-------|------------|---------|
| Train | 56,484     | ~390 MB |
| Test  | 1,000      | ~7 MB   |

## Usage

Install the 🤗 Datasets library:

```bash
pip install datasets
```

Load the dataset:

```python
from datasets import load_dataset

dataset = load_dataset("ming030890/youtube_caption_yue")
train_data = dataset["train"]
print(train_data[0])
```

Play audio (in notebooks):

```python
from IPython.display import Audio
example = train_data[0]
Audio(example["audio"]["array"], rate=example["audio"]["sampling_rate"])
```

## License

Free for research use. Check original YouTube licenses before reuse or redistribution.