brendayy/lrlid

Dataset Card for Dataset Name

Status: Design-stage dataset card (pre-annotation, pre-evaluation)

Dataset Card for AAPB Low-Resource Language Identification Dataset (v0)

This dataset contains archival broadcast audio from the American Archive of Public Broadcasting (AAPB), curated for evaluating language identification (LID) systems on low-resource languages in realistic archival conditions. The initial focus is on Samoan, Yupik, and Spanish.

This dataset card aims to be a base template for new datasets. It has been generated using this raw template.

Dataset Details

Dataset Description

The AAPB Low-Resource Language Identification Dataset is a collection of real-world broadcast audio drawn from the American Archive of Public Broadcasting (AAPB). The dataset is designed to support evaluation of existing off-the-shelf language identification models on low-resource languages under challenging archival conditions, including code-switching, non-speech segments, and variable audio quality.

This dataset card represents a design-stage (v0) description. Annotation and evaluation are ongoing, and quantitative results will be added iteratively in future versions.

• Curated by: Brandeis University, in collaboration with GBH and the CLAMS project
• Funded by [optional]: [More Information Needed]
• Shared by [optional]: [More Information Needed]
• Language(s) (NLP): [More Information Needed]
• License: [More Information Needed]

Dataset Sources [optional]

• Repository: [TBD – internal CLAMS / Brandeis repositories]
• Paper [optional]: [More Information Needed]
• Demo [optional]: [More Information Needed]

Uses

Direct Use

This dataset is intended for:

• Evaluation of existing language identification (LID) models on selected low-resource languages
• Analysis of LID model behavior on archival broadcast audio
• Studying the effects of code-switching, non-speech segments, and acoustic variability on LID output
• Integration with the CLAMS evaluation framework for reproducible benchmarkin

[More Information Needed]

Out-of-Scope Use

This dataset is not intended for:

• Training new language identification models
• Speaker identification or speaker profiling --?
• Commercial deployment or monetization
• Applications requiring fine-grained speaker attribution or demographic inference --?

[More Information Needed]

Dataset Structure

The dataset consists of audio items corresponding to AAPB broadcast programs or program segments. Each item may contain multiple speakers, multiple languages, and non-speech content such as music.

Annotations, where present, are stored in [?]format and represent time-aligned language labels for speech segments.

At the current stage, the dataset is not split into train/validation/test partitions, as it is primarily intended for evaluation rather than training.

[More Information Needed]

Dataset Creation

Curation Rationale

This dataset was created to address the lack of benchmarks for language identification on low-resource languages in realistic archival contexts. Many existing datasets focus on clean or synthetic speech, which does not reflect the conditions encountered in broadcast archives.

By using authentic AAPB material, this dataset aims to surface practical failure modes and limitations of current LID systems.

[More Information Needed]

Source Data

Data Collection and Processing

Audio data was selected from AAPB holdings based on:

• Likely presence of the target language(s)
• Sufficient duration for language identification analysis
• Representation of real broadcast conditions
• (Minimal normalization is applied in order to preserve the acoustic characteristics of archival audio).

[More Information Needed]

Who are the source data producers?

The source data was produced by human speakers participating in broadcast programs, including hosts, interviewees, and callers. Demographic or identity information about speakers is generally unknown and is not inferred. The data was originally created for broadcast and archival purposes, not for machine learning.

[More Information Needed]

Annotations [optional]

Annotation process

Annotation focuses on identifying the language of speech segments relevant for LID evaluation. Annotations are being introduced incrementally and may be refined based on pilot evaluation results. Annotation guidelines are under development and may evolve as new error patterns are observed. [More Information Needed]

Who are the annotators?

Annotations are produced by human annotators fluent or native in the target languages, including (summer) fellows recruited for Spanish, Samoan, and Yupik. Annotators are selected based on language proficiency.

[More Information Needed]

Personal and Sensitive Information

The dataset consists of publicly archived broadcast material. No additional personal or sensitive information is intentionally added during annotation. Speakers may be identifiable as part of the original broadcast context, but no attempts are made to infer or augment personal attributes.

[More Information Needed]

Bias, Risks, and Limitations

This dataset has several known limitations:

• Small sample sizes for low-resource languages
• Language imbalance and frequent code-switching with English
• Presence of non-speech content such as music and background noise
• Variability in recording quality across archival programs
• Potential ambiguity at segment boundaries during annotation

These limitations should be considered when interpreting evaluation results.

[More Information Needed]

Recommendations

Users should be made aware of the risks, biases and limitations of the dataset. More information needed for further recommendations.

Users are encouraged to:

• Treat evaluation results as indicative rather than definitive
• Avoid overgeneralizing findings beyond archival broadcast contexts
• Report observed biases or failure modes in downstream analyses

Citation [optional]

BibTeX:

[More Information Needed]

APA:

[More Information Needed]

Glossary [optional]

[More Information Needed]

More Information [optional]

[More Information Needed]

Dataset Card Authors [optional]

[More Information Needed]

Dataset Card Contact

Yangyang Chen (Brandeis University), with contributions from GBH and CLAMS collaborators

[More Information Needed]