wav2vec2-large-xls-r-300m-odia

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

python eval.py --model_id ./ --dataset mozilla-foundation/common_voice_7_0 --config as --split test --log_outputs

• WER: 1.0921052631578947
• CER: 2.5547945205479454

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

Training machine details

• Platform: Linux-5.11.0-37-generic-x86_64-with-glibc2.10
• CPU cores: 60
• Python version: 3.8.8
• PyTorch version: 1.10.1+cu102
• GPU is visible: True
• Transformers version: 4.16.0.dev0
• Datasets version: 1.17.1.dev0
• soundfile version: 0.10.3

Training script

python run_speech_recognition_ctc.py \
    --dataset_name="mozilla-foundation/common_voice_7_0" \
    --model_name_or_path="facebook/wav2vec2-xls-r-300m" \
    --dataset_config_name="or" \
    --output_dir="./wav2vec2-large-xls-r-300m-odia" \
    --overwrite_output_dir \
    --num_train_epochs="120" \
    --per_device_train_batch_size="16" \
    --per_device_eval_batch_size="16" \
    --gradient_accumulation_steps="2" \
    --learning_rate="7.5e-5" \
    --warmup_steps="500" \
    --length_column_name="input_length" \
    --evaluation_strategy="steps" \
    --text_column_name="sentence" \
    --chars_to_ignore , ? . ! \- \; \: \" “ % ‘ ” � — \’ … \– \' \’ \– \
    --save_steps="500" \
    --eval_steps="500" \
    --logging_steps="100" \
    --layerdrop="0.0" \
    --activation_dropout="0.1" \
    --save_total_limit="3" \
    --freeze_feature_encoder \
    --feat_proj_dropout="0.0" \
    --mask_time_prob="0.75" \
    --mask_time_length="10" \
    --mask_feature_prob="0.25" \
    --mask_feature_length="64" \
    --gradient_checkpointing \
    --use_auth_token \
    --fp16 \
    --group_by_length \
    --do_train --do_eval \
  --push_to_hub

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 7.5e-05
• train_batch_size: 16
• eval_batch_size: 16
• 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: 120.0
• mixed_precision_training: Native AMP

Training results

	eval_loss	eval_wer	eval_runtime	eval_samples_per_second	eval_steps_per_second	epoch
0	3.35224	0.998972	5.0475	22.189	1.387	29.41
1	1.33679	0.938335	5.0633	22.12	1.382	58.82
2	0.737202	0.957862	5.0913	21.998	1.375	88.24
3	0.658212	0.96814	5.0953	21.981	1.374	117.65
4	0.658	0.9712	5.0953	22.115	1.382	120

Framework versions

• Transformers 4.16.0.dev0
• Pytorch 1.10.1+cu102
• Datasets 1.17.1.dev0
• Tokenizers 0.11.0