lucio/xls-r-uyghur-cv7

XLS-R-300M Uyghur CV7

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - UG dataset. It achieves the following results on the evaluation set:

• Loss: 0.1772
• Wer: 0.2589

Model description

For a description of the model architecture, see facebook/wav2vec2-xls-r-300m

The model vocabulary consists of the alphabetic characters of the Perso-Arabic script for the Uyghur language, with punctuation removed.

Intended uses & limitations

This model is expected to be of some utility for low-fidelity use cases such as:

• Draft video captions
• Indexing of recorded broadcasts

The model is not reliable enough to use as a substitute for live captions for accessibility purposes, and it should not be used in a manner that would infringe the privacy of any of the contributors to the Common Voice dataset nor any other speakers.

Training and evaluation data

The combination of train and dev of common voice official splits were used as training data. The official test split was used as validation data as well as for final evaluation.

Training procedure

The featurization layers of the XLS-R model are frozen while tuning a final CTC/LM layer on the Uyghur CV7 example sentences. A ramped learning rate is used with an initial warmup phase of 2000 steps, a max of 0.0001, and cooling back towards 0 for the remainder of the 18500 steps (100 epochs).

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2000
• num_epochs: 100.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.3043	2.73	500	3.2415	1.0
3.0482	5.46	1000	2.9591	1.0
1.4767	8.2	1500	0.4779	0.5777
1.3152	10.93	2000	0.3697	0.4938
1.2246	13.66	2500	0.3084	0.4459
1.1781	16.39	3000	0.2842	0.4154
1.1351	19.13	3500	0.2615	0.3929
1.1052	21.86	4000	0.2462	0.3747
1.0711	24.59	4500	0.2366	0.3652
1.035	27.32	5000	0.2268	0.3557
1.0277	30.05	5500	0.2243	0.3450
1.002	32.79	6000	0.2204	0.3389
0.9837	35.52	6500	0.2156	0.3349
0.9773	38.25	7000	0.2127	0.3289
0.9807	40.98	7500	0.2142	0.3274
0.9582	43.72	8000	0.2004	0.3142
0.9548	46.45	8500	0.2022	0.3050
0.9251	49.18	9000	0.2019	0.3035
0.9103	51.91	9500	0.1964	0.3021
0.915	54.64	10000	0.1970	0.3032
0.8962	57.38	10500	0.2007	0.3046
0.8729	60.11	11000	0.1967	0.2942
0.8744	62.84	11500	0.1952	0.2885
0.874	65.57	12000	0.1894	0.2895
0.8457	68.31	12500	0.1895	0.2828
0.8519	71.04	13000	0.1912	0.2875
0.8301	73.77	13500	0.1878	0.2760
0.8226	76.5	14000	0.1808	0.2701
0.8071	79.23	14500	0.1849	0.2741
0.7999	81.97	15000	0.1808	0.2717
0.7947	84.7	15500	0.1821	0.2716
0.7783	87.43	16000	0.1824	0.2661
0.7729	90.16	16500	0.1773	0.2639
0.7759	92.9	17000	0.1767	0.2629
0.7713	95.63	17500	0.1780	0.2621
0.7628	98.36	18000	0.1773	0.2594

Framework versions

• Transformers 4.16.0.dev0
• Pytorch 1.10.1+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0