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… See the full description on the dataset page: https://huggingface.co/datasets/LEMAS-Project/LEMAS-Dataset-train.
906.5k
1---2license: cc-by-4.03language:4- it5- pt6- es7- fr8- de9- vi10- id11- ru12- en13- zh14task_categories:15- text-to-speech16- automatic-speech-recognition17size_categories:18- n>1T19---20 21## Overview22 23This dataset is part of **LEMAS-Project** ([lemas-project.github.io/LEMAS-Project](https://lemas-project.github.io/LEMAS-Project/)).24It contains a large-scale training set (150k+ hours) and a curated evaluation set25(500 utterances per language) covering 10 languages, all with word-level alignment.26 27 28## Fields29- `key`: unique utterance identifier; the first two characters indicate the language ID30- `audio`: relative path to the MP3 audio file (in the eval set, this key is renamed to "file_name" for compatibility with the viewer)31- `dur`: audio duration in seconds32- `txt`: original transcription33- `align`: alignment information, including:34 - `align.txt`: normalized text used for alignment35 - `align.words`: list of word-level timestamps and confidence scores36 37 38 39## Methods40 41### Train Set42 43- The training set is constructed by filtering large-scale aligned audio–text pairs with language- and dataset-specific constraints.44- URL: [https://huggingface.co/datasets/LEMAS-Project/LEMAS-Dataset-train](https://huggingface.co/datasets/LEMAS-Project/LEMAS-Dataset-train)45 46**Filtering rules:**47- Only samples with successfully extracted word-level alignments are kept; failed alignments are skipped.48- The average alignment confidence score is greater than a threshold **x**, where **x ∈ [0.2, 0.5]** depending on the source dataset.49- Audio duration is between **0.5 and 30 seconds**, and the maximum pause between consecutive words does not exceed **4 seconds**.50- Text is restricted to supported languages; samples containing characters outside the supported set (10 languages) are removed.51- The text-to-duration ratio `len(txt) / dur` falls within a language-specific range.52 53 54#### **Statistics**55| lang | utterances | total_dur(h) | avg_dur(s) | total_chars | avg_chars | char/sec | total_words | avg_words | word/sec |56|------|-----------:|-------------:|-----------:|------------:|----------:|---------:|------------:|----------:|---------:|57| it | 7209567 | 6116.69 | 3.054 | 281128388 | 38.994 | 12.7669 | 48859270 | 6.777 | 2.2188 |58| fr | 7557428 | 6535.94 | 3.113 | 344633332 | 45.602 | 14.6469 | 65843169 | 8.712 | 2.7983 |59| vi | 6426292 | 6600.07 | 3.697 | 387621747 | 60.318 | 16.3139 | 89903814 | 13.990 | 3.7838 |60| pt | 8682442 | 7384.42 | 3.062 | 339417251 | 39.092 | 12.7678 | 60316112 | 6.947 | 2.2689 |61| de | 11003833 | 9842.14 | 3.220 | 487581808 | 44.310 | 13.7612 | 80119609 | 7.281 | 2.2612 |62| id | 11659550 | 11246.27 | 3.472 | 543857918 | 46.645 | 13.4331 | 85967394 | 7.373 | 2.1234 |63| es | 26407271 | 21224.55 | 2.893 | 1011116926 | 38.289 | 13.2331 | 183673862 | 6.955 | 2.4038 |64| ru | 27474400 | 22919.31 | 3.003 | 991530233 | 36.089 | 12.0172 | 163018329 | 5.933 | 1.9758 |65| en | 9515267 | 25347.90 | 9.590 | 1419864294 | 149.220 | 15.5597 | 268676221 | 28.236 | 2.9443 |66| zh | 17776663 | 32919.28 | 6.667 | 2474249286 | 139.185 | 20.8781 | 496957308 | 27.956 | 4.1934 |67 68Words and chars statistics are computed based on the normalized alignment text (`align.txt`).69 70### Eval Set71 72- The eval set is built by filtering, trimming, and ranking aligned audio–text pairs.73- URL: [https://huggingface.co/datasets/LEMAS-Project/LEMAS-Dataset-eval](https://huggingface.co/datasets/LEMAS-Project/LEMAS-Dataset-eval)74 75**Filtering rules:**76- Average word-level alignment score > **0.9**77- Number of aligned words > **5**78- Duration between **3 and 15 seconds**79- Sentence-end silence is trimmed to at most **0.2s**80 81**Selection:**82- Samples are ranked by 83 `final_score = edge_gap × density_diff`, where 84 `edge_gap = words[0].start + (dur - words[-1].end)` and 85 `density_diff = |len(align_txt)/dur − global_mean_density|`86 87#### **Statistics**88 89| lang | utterances | total_dur(min) | avg_dur(s) | total_chars | avg_chars | char/sec | total_words | avg_words | word/sec |90|------|-----------:|---------------:|-----------:|------------:|----------:|---------:|------------:|----------:|---------:|91| it | 500 | 44.22 | 5.306 | 40388 | 80.78 | 15.22 | 6599 | 13.20 | 2.49 |92| fr | 500 | 38.17 | 4.580 | 38098 | 76.20 | 16.64 | 6546 | 13.09 | 2.86 |93| vi | 500 | 36.74 | 4.409 | 28546 | 57.09 | 12.95 | 6727 | 13.45 | 3.05 |94| pt | 500 | 41.69 | 5.003 | 33343 | 66.69 | 13.33 | 5812 | 11.62 | 2.32 |95| de | 500 | 38.65 | 4.638 | 36571 | 73.14 | 15.77 | 5599 | 11.20 | 2.41 |96| id | 500 | 47.20 | 5.665 | 41026 | 82.05 | 14.49 | 6133 | 12.27 | 2.17 |97| es | 500 | 40.52 | 4.862 | 37075 | 74.15 | 15.25 | 6216 | 12.43 | 2.56 |98| ru | 500 | 40.24 | 4.828 | 33886 | 67.77 | 14.04 | 5138 | 10.28 | 2.13 |99| en | 500 | 67.46 | 8.095 | 62449 | 124.90 | 15.43 | 11325 | 22.65 | 2.80 |100| zh | 500 | 75.84 | 9.101 | 95669 | 191.34 | 21.02 | 18627 | 37.25 | 4.09 |101 102Statistics are computed based on trimmed audio and normalized alignment text (`align.txt`).103 104 105## Citation106[https://arxiv.org/abs/2601.04233](https://arxiv.org/abs/2601.04233)107```108@article{zhao2026lemas,109 title={LEMAS: A 150K-Hour Large-scale Extensible Multilingual Audio Suite with Generative Speech Models},110 author={Zhao, Zhiyuan and Lin, Lijian and Zhu, Ye and Xie, Kai and Liu, Yunfei and Li, Yu},111 journal={arXiv preprint arXiv:2601.04233},112 year={2026}113}114```