andrewbawitlung/wav2vec2-xls-r-2b-mizo-lus-v25

Mizo Automatic Speech Recognition

This model is a fine-tuned version of facebook/wav2vec2-xls-r-2b on the generator dataset. It achieves the following results on the evaluation set:

• Loss: 0.1318
• Wer: 0.1582

Citation

BibTeX entry and citation info:

@article{10.1145/3746063,
author = {Bawitlung, Andrew and Dash, Sandeep Kumar and Pattanayak, Radha Mohan},
title = {Mizo Automatic Speech Recognition: Leveraging Wav2vec 2.0 and XLS-R for Enhanced Accuracy in Low-Resource Language Processing},
year = {2025},
url = {https://doi.org/10.1145/3746063},
doi = {10.1145/3746063},
journal = {ACM Trans. Asian Low-Resour. Lang. Inf. Process.},
month = jun,
}

Training and evaluation data

MiZonal v1.0

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 8
• eval_batch_size: 8
• seed: 49
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 28
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.1831	2.18	300	0.3039	0.3949
0.7933	4.35	600	0.1978	0.2774
0.5182	6.53	900	0.1940	0.2764
0.4319	8.7	1200	0.1549	0.2587
0.3472	10.88	1500	0.1345	0.2096
0.2858	13.06	1800	0.1374	0.1916
0.2349	15.23	2100	0.1298	0.1839
0.1973	17.41	2400	0.1214	0.1820
0.1633	19.58	2700	0.1302	0.1631
0.1378	21.76	3000	0.1331	0.1804
0.1131	23.93	3300	0.1284	0.1666
0.0905	26.11	3600	0.1318	0.1582

Framework versions

• Transformers 4.37.2
• Pytorch 2.3.1+cu121
• Datasets 2.16.1
• Tokenizers 0.15.1