---
language:
- uk
license: other
task_categories:
- automatic-speech-recognition
pretty_name: Toronto TV Ukrainian Speech Dataset
size_categories:
- 10K<n<100K
tags:
- audio
- speech
- ukrainian
- asr
dataset_info:
  features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 44100
  - name: text
    dtype: string
  - name: video_id
    dtype: string
  splits:
  - name: train
    num_bytes: 9621946369
    num_examples: 18303
  download_size: 9621213392
  dataset_size: 9621946369
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# Toronto TV Ukrainian Speech Dataset

> **For educational purposes only.**
> All rights to the original video and audio content belong to **Телебачення Торонто** ([YouTube channel](https://www.youtube.com/channel/UCF_ZiWz2Vcq1o5u5i1TT3Kw)).

---

## Dataset Summary

A Ukrainian-language speech dataset parsed from the [Телебачення Торонто](https://www.youtube.com/channel/UCF_ZiWz2Vcq1o5u5i1TT3Kw) YouTube channel. Each sample consists of a short audio clip and its corresponding Ukrainian subtitle text, intended for use in automatic speech recognition (ASR) research and education.

## Dataset Details

| Property | Value |
|---|---|
| Language | Ukrainian (`uk`) |
| Task | Automatic Speech Recognition (ASR) |
| Total audio duration | ~30.3 hours |
| Total clips | 18,303 |
| Source videos | 71 |
| Average clip duration | ~6 seconds |
| Audio format | WAV, 44.1 kHz |
| Text format | Ukrainian subtitles |

## Dataset Structure

```
toronto-dataset/
├── labels.jsonl                  # Maps audio paths to transcriptions
├── toronto_0/
│   ├── toronto_0_0.wav
│   ├── toronto_0_1.wav
│   └── ...
├── toronto_1/
│   └── ...
└── ...                           # 71 folders total (one per source video)
```

Each folder corresponds to one source video. Files within a folder are named `{video_id}_{segment_index}.wav` and are ordered chronologically.

### `labels.jsonl`

A single JSON object mapping relative audio paths to their transcription strings:

```json
{
  "dataset/toronto_0/toronto_0_0.wav": "Слава Ісу! ...",
  "dataset/toronto_0/toronto_0_1.wav": "...",
  ...
}
```

### Data Fields

- `audio` — WAV audio clip (short utterance, ~6 sec average)
- `text` — Ukrainian transcription from the original subtitles
- `video_id` — source video folder name (e.g. `toronto_0`), useful for train/test splitting by video

## Usage (HuggingFace `datasets`)

```python
from datasets import load_dataset

ds = load_dataset("yuriilaba/toronto-tv-ukrainian")
print(ds["train"][0])
# {'audio': {'array': [...], 'sampling_rate': 16000}, 'text': 'Слава Ісу! ...'}
```

## License & Attribution

This dataset was created **for educational and research purposes only**.

- **Original content**: All rights belong to [Телебачення Торонто](https://www.youtube.com/channel/UCF_ZiWz2Vcq1o5u5i1TT3Kw).
- **Dataset created by**:
  - [Yurii Laba](https://www.linkedin.com/in/yura-laba/)
  - [Volodymyr Mudriy](https://www.linkedin.com/in/volodymyr-mudryi/)

If you are the rights holder and have concerns about this dataset, please contact the authors.