ivangtorre/second_americas_nlp_2022

Second AmericasNLP 2022

Dataset Summary

This dataset contains the speech data released as part of the Second Workshop on Natural Language Processing for Indigenous Languages of the Americas (AmericasNLP 2022). It provides audio recordings and corresponding transcriptions for several Indigenous languages of the Americas together with Spanish, supporting research on multilingual and low-resource Automatic Speech Recognition (ASR).

This Hugging Face version has been prepared to facilitate reproducible research and straightforward integration with the datasets library.

The dataset is intended for:

• Automatic Speech Recognition (ASR)
• Self-supervised speech representation learning
• Speech encoder evaluation
• Cross-lingual transfer learning
• Low-resource speech research
• Benchmarking multilingual speech models

---

Supported Languages

The dataset contains recordings from the following languages:

Language	ISO Code
Spanish	es
Quechua	qu
Aymara	ay
Guaraní	gn
Nahuatl	nah

---

Dataset Structure

Each example contains an audio recording together with its associated transcription and metadata.

Typical fields include:

Field	Description
audio	Audio waveform loaded by Hugging Face
transcription	Ground-truth transcription
language	Language identifier
speaker_id	Speaker identifier (when available)
split	Dataset split (train/validation/test)

Depending on the original source files, additional metadata may also be available.

---

Audio Format

The dataset stores recordings in standard WAV format.

When loaded through the Hugging Face datasets library, the audio column is automatically decoded as:

{
    "array": np.ndarray,
    "sampling_rate": int,
    "path": str
}

Example:

from datasets import load_dataset

dataset = load_dataset("ivangtorre/second_americas_nlp_2022")

sample = dataset["train"][0]

print(sample["audio"]["sampling_rate"])
print(sample["transcription"])

---

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("ivangtorre/second_americas_nlp_2022")

Access an audio sample:

sample = dataset["train"][0]

audio = sample["audio"]["array"]
sampling_rate = sample["audio"]["sampling_rate"]
text = sample["transcription"]

---

Source Data

The original dataset was released as part of the Second Workshop on Natural Language Processing for Indigenous Languages of the Americas (AmericasNLP 2022).

This repository republishes the dataset in Hugging Face format for easier access and reproducible experimentation.

No modifications have been made to the original speech recordings or transcriptions beyond adapting the data structure for compatibility with the Hugging Face datasets library.

---

Acknowledgements

We gratefully acknowledge the organizers of the Second AmericasNLP Workshop and all contributors involved in collecting, curating, and releasing the original speech corpus.

This Hugging Face version was prepared by Iván G. Torre to improve accessibility, interoperability, and reproducibility for the speech research community.

---

Citation

If you use this dataset, please cite the original AmericasNLP 2022 dataset and workshop publication.

@inproceedings{ebrahimi2022findings,
  title={Findings of the Second AmericasNLP Competition on Speech-to-Text Translation},
  author={Ebrahimi, Abteen and Mager, Manuel and Wiemerslage, Adam and Denisov, Pavel and Oncevay, Arturo and Liu, Danni and Koneru, Sai and Ugan, Enes Yavuz and Li, Zhaolin and Niehues, Jan and others},
  booktitle={NeurIPS 2022 Competition Track},
  pages={217--232},
  year={2022},
  organization={PMLR}
}