By clicking 'I agree', you agree to the following: This corpus is provided under the Attribution-NonCommercial-ShareAlike 3.0 Unported license. I will use this corpus for the purposes of the IWSLT 2026 Low-Resource Shared Task, for research purposes only. I will not redistribute the data under a different license or use it for commercial purposes.
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Log in or Sign Up to review the conditions and access this dataset content.
Mapuzugun to Spanish Data for IWSLT 2026 Low-resource ST track
This is the speech-to-text translation data in Mapuzugun to Spanish for the IWSLT 2026 Low-resource ST track.
Mapudungun is a language isolate and the indigenous language of the Mapuche people, spoken by about 100,000-200,000 people in Chile and Argentina.
Data are based on the corpus described in this paper and this Github repository.
Currently, we release the train and dev sets. Later, we will release the test audios here.
Data Details
Example usage:
from datasets import load_dataset
train_data = load_dataset("mengct00/Mapudungun_iwslt26")["train"]
for example in train_data:
audio = example["audio"]
audio_array = audio["array"] # shape: (T, )
sampling_rate = audio["sampling_rate"]
assert sampling_rate == 44100
transcript_raw = example["arn"]
transcript_clean = example["arn_clean"]
translation = example["spa"]
Data Fields:
audio: audio, with a sampling rate 44100.arn: raw transcriptions.arn_clean: cleaned transcriptions. We (1) applied functions from the original repo 1 2; (2) removed all brackets; (3) standardized spaces.spa: translation.
Note: The final ranking will only be based on translations not transcriptions. You may apply any methods to clean the transcriptions, if you think it's needed.
Statistics:
| Split | #Utterances | #Hours |
|---|---|---|
| Train | 41092 | 78.504 |
| Validation | 1201 | 2.438 |
| Test | 7411 | 13.558 |
- Downloads last month
- 69