LEMAS-Project/LEMAS-Dataset-train

Overview

This dataset is part of LEMAS-Project (lemas-project.github.io/LEMAS-Project). It contains a large-scale training set (150k+ hours) and a curated evaluation set (500 utterances per language) covering 10 languages, all with word-level alignment.

Fields

• —key: unique utterance identifier; the first two characters indicate the language ID
• —audio: relative path to the MP3 audio file (in the eval set, this key is renamed to "file_name" for compatibility with the viewer)
• —dur: audio duration in seconds
• —txt: original transcription
• —align: alignment information, including:
• —align.txt: normalized text used for alignment
• —align.words: list of word-level timestamps and confidence scores

Methods

Train Set

• —The training set is constructed by filtering large-scale aligned audio–text pairs with language- and dataset-specific constraints.
• —URL: https://huggingface.co/datasets/LEMAS-Project/LEMAS-Dataset-train

Filtering rules:

• —Only samples with successfully extracted word-level alignments are kept; failed alignments are skipped.
• —The average alignment confidence score is greater than a threshold x, where x ∈ [0.2, 0.5] depending on the source dataset.
• —Audio duration is between 0.5 and 30 seconds, and the maximum pause between consecutive words does not exceed 4 seconds.
• —Text is restricted to supported languages; samples containing characters outside the supported set (10 languages) are removed.
• —The text-to-duration ratio len(txt) / dur falls within a language-specific range.

Statistics

Words and chars statistics are computed based on the normalized alignment text (align.txt).

Eval Set

• —The eval set is built by filtering, trimming, and ranking aligned audio–text pairs.
• —URL: https://huggingface.co/datasets/LEMAS-Project/LEMAS-Dataset-eval

Filtering rules:

• —Average word-level alignment score > 0.9
• —Number of aligned words > 5
• —Duration between 3 and 15 seconds
• —Sentence-end silence is trimmed to at most 0.2s

Selection:

• —Samples are ranked by final_score = edge_gap × density_diff, where edge_gap = words[0].start + (dur - words[-1].end) and density_diff = |len(align_txt)/dur − global_mean_density|

Statistics

Statistics are computed based on trimmed audio and normalized alignment text (align.txt).

Citation

https://arxiv.org/abs/2601.04233

@article{zhao2026lemas,
  title={LEMAS: A 150K-Hour Large-scale Extensible Multilingual Audio Suite with Generative Speech Models},
  author={Zhao, Zhiyuan and Lin, Lijian and Zhu, Ye and Xie, Kai and Liu, Yunfei and Li, Yu},
  journal={arXiv preprint arXiv:2601.04233},
  year={2026}
}