---
dataset_info:
  features:
  - name: audio
    dtype: audio
  - name: text_en
    dtype: string
  - name: text_mn
    dtype: string
  - name: filename
    dtype: string
  - name: group
    dtype: string
  - name: audio_language
    dtype: string
  splits:
  - name: en
    num_bytes: 6332410145.572
    num_examples: 5244
  - name: mn
    num_bytes: 499503079
    num_examples: 417
  download_size: 6811164844
  dataset_size: 6831913224.572
configs:
- config_name: default
  data_files:
  - split: en
    path: data/en-*
  - split: mn
    path: data/mn-*
task_categories:
- automatic-speech-recognition
- translation
language:
- en
- mn
tags:
- audio
---

# TED & TEDx Parallel Corpus (English-Mongolian)

The dataset is composed of two distinct subsets:

- TED Talks (split `en`): English-language talks sourced from the official TED platform, paired with high-quality, human-generated Mongolian subtitles.
- TEDxUlaanbaatar (split `mn`): Mongolian-language talks from local TEDx events in Ulaanbaatar, paired with the original Mongolian subtitles and machine-translated English subtitles.

This version of the dataset features segmented audio and text, with each segment having a maximum duration of 30 seconds. For the complete, unsegmented version, please refer to [bilguun/ted_talks_en_mn](https://huggingface.co/datasets/bilguun/ted_talks_en_mn).

## Known Limitations

Machine Translation Quality: The primary limitation is the quality of the English translations in the TEDxUlaanbaatar split. As these are machine-generated, they may contain inaccuracies, grammatical errors, or mistranslations of nuanced or idiomatic expressions.

Subtitle Alignment and Errors: As the data is derived from subtitles, some entries may contain minor errors. This can include missing words or phrases, or slight mismatches between parallel sentences due to subtitle timing and segmentation. Users should consider a preprocessing step to handle potential misalignments.