wav2vec2-large-xls-r-300m-ia

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

• Loss: 0.1452
• Wer: 0.1253

Training Procedure

Training is conducted in Google Colab, the training notebook provided in the repo

Training and evaluation data

Language Model Created from texts from processed sentence in train + validation split of dataset (common voice 8.0 for Interlingua) Evaluation is conducted in Notebook, you can see within the repo "notebook_evaluation_wav2vec2_ia.ipynb"

Test WER without LM wer = 20.1776 % cer = 4.7205 %

Test WER using wer = 8.6074 % cer = 2.4147 %

evaluation using eval.py

huggingface-cli login #login to huggingface for getting auth token to access the common voice v8
#running with LM
python eval.py --model_id ayameRushia/wav2vec2-large-xls-r-300m-ia --dataset mozilla-foundation/common_voice_8_0 --config ia --split test

# running without LM
python eval.py --model_id ayameRushia/wav2vec2-large-xls-r-300m-ia --dataset mozilla-foundation/common_voice_8_0 --config ia --split test --greedy

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 16
• eval_batch_size: 4
• 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: 400
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
7.432	1.87	400	2.9636	1.0
2.6922	3.74	800	2.2111	0.9977
1.2581	5.61	1200	0.4864	0.4028
0.6232	7.48	1600	0.2807	0.2413
0.4479	9.35	2000	0.2219	0.1885
0.3654	11.21	2400	0.1886	0.1606
0.323	13.08	2800	0.1716	0.1444
0.2935	14.95	3200	0.1687	0.1443
0.2707	16.82	3600	0.1632	0.1382
0.2559	18.69	4000	0.1507	0.1337
0.2433	20.56	4400	0.1572	0.1358
0.2338	22.43	4800	0.1489	0.1305
0.2258	24.3	5200	0.1485	0.1278
0.2218	26.17	5600	0.1470	0.1272
0.2169	28.04	6000	0.1470	0.1270
0.2117	29.91	6400	0.1452	0.1253

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.0+cu111
• Datasets 1.18.3
• Tokenizers 0.11.0