---
language: fa
pretty_name: Farsi Youtube 2024 ASR Dataset
tags:
- Farsi
- Persian
- ASR
- youtube
task_categories:
  - automatic-speech-recognition
dataset_size: "N > 400k"
dataset_info:
 splits:
  - name: unvalidated
    num_examples: 425468
license: cc0-1.0
---

# Farsi Youtube 2024 ASR Dataset

This dataset consists of over **385** hours of transcribed audio extracted from various YouTube videos in the Persian language (more than 400k rows).

## Dataset Description

The dataset includes Farsi content from various types of videos spanning from older productions up to mid-2024, including:
- Podcasts
- TV Shows
- Educational Content
- Interviews
- Documentaries

Utterances and sentences are extracted based on the timing of subtitles.

The list of videos used in this dataset is stored in the `yt_ids.csv` file as follows:
```
13XpMM7RT2c	20231207	سرگذشت پُل پوت هیولای کامبوج و رهبر خمرهای سرخ
yU6LtnpVKLo	20231210	راز بزرگترین جاسوس عرب|بیوگرافی اشرف مروان
b9cTFkO6Q18	20231214	دقیقا چه اتفاقی افتاده؟ بالاخره توی این درگیری کی پیروز شد؟ 7 -27 نوامبر
wW76xHcxw48	20231217	حقایق شنیده نشده درباره نجات دنیا از جنگ هسته ای!
pr1dNDD6viM	20231123	افشای زندگی صدام حسین! | قسمت دوم
...
```

## Note
This dataset contains raw, unvalidated auto-generated transcriptions. Transcriptions may include inaccuracies due to mal-transcriptions, and timing may occasionally be imprecise. Many efforts have been made to cleanse the data using various methods and software.
Users are advised to:
- Perform their own quality assessment
- Create their own train/validation/test splits based on their specific needs
- Validate a subset of the data if needed for their use case

For validating the data you can use [AnnoTitan](https://github.com/dhpour/annotitan), which is a crowdsourcing app developed for such ASR data.

## Usage
<details>

Huggingface datasets library:
```python
from datasets import load_dataset
dataset = load_dataset('PerSets/youtube-persian-asr', trust_remote_code=True)
```
</details>