wav2vec2-telugu_150

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the openslr dataset.

It achieves the following results on the evaluation set:

• Loss: 0.3312
• Wer: 0.2213

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• 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: 500
• num_epochs: 150
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
6.096	3.84	400	0.5762	0.7029
0.427	7.69	800	0.3124	0.5148
0.208	11.54	1200	0.2994	0.4201
0.1506	15.38	1600	0.3106	0.3844
0.1223	19.23	2000	0.3080	0.3608
0.1094	23.08	2400	0.3206	0.3332
0.0949	26.92	2800	0.3085	0.3253
0.0802	30.77	3200	0.3076	0.3425
0.0713	34.61	3600	0.3280	0.3398
0.0687	38.46	4000	0.3042	0.3081
0.0613	42.31	4400	0.3227	0.3073
0.0548	46.15	4800	0.3152	0.3213
0.0508	50.0	5200	0.3259	0.3107
0.0455	53.84	5600	0.3046	0.2881
0.0427	57.69	6000	0.2779	0.3007
0.0391	61.54	6400	0.2996	0.2693
0.0388	65.38	6800	0.3016	0.2695
0.0339	69.23	7200	0.3225	0.2935
0.0312	73.08	7600	0.2907	0.2942
0.029	76.92	8000	0.3148	0.3029
0.0254	80.77	8400	0.3118	0.2996
0.0229	84.61	8800	0.3022	0.2993
0.0231	88.46	9200	0.3203	0.2465
0.019	92.31	9600	0.3223	0.2460
0.0173	96.15	10000	0.3178	0.2501
0.0168	100.0	10400	0.2937	0.2415
0.015	103.84	10800	0.3062	0.2415
0.014	107.69	11200	0.3104	0.2383
0.012	111.54	11600	0.3308	0.2408
0.0111	115.38	12000	0.3228	0.2335
0.01	119.23	12400	0.3228	0.2374
0.0096	123.08	12800	0.3241	0.2304
0.009	126.92	13200	0.3237	0.2295
0.0075	130.77	13600	0.3221	0.2261
0.0065	134.61	14000	0.3310	0.2277
0.0064	138.46	14400	0.3348	0.2266
0.0064	142.31	14800	0.3330	0.2229
0.0056	146.15	15200	0.3310	0.2229
0.0053	150.0	15600	0.3312	0.2213

Test results

WER(without LM): 42.8%

WER(with LM): 42%

Framework versions

• Transformers 4.24.0
• Pytorch 1.13.0+cu117
• Datasets 2.6.1
• Tokenizers 0.13.2

ps: 150 in repo name denotes number of epochs