wav2vec2-large-xlsr-300m-nepali

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m

The dataset used to train are :

1. OpenSLR-54 Corpus
2. External Data

For evaluation on publicly available datasets, can use OpenSLR-43 corpus - Model is not trained on this data which is also available on HuggingFace as train-set where it achieves 27% WER and 8.3% CER with 5 gram language model.

Script to Evaluate the Model on OpenSLR-43 train set :

python3 eval.py --model_id prajin/wav2vec2-large-xlsr-300m-nepali --dataset openslr --config SLR43  --split train --log_outputs

Below evaluation result is the evaluation on separated 10000 samples from total training dataset.

It achieves the following results on the evaluation set Without using Language Model :

• Loss: 0.2625
• Wer: 0.3426

With Language model ( 5 gram )

• Wer: 0.2502

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 6e-05
• 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: 1000
• num_epochs: 1
• mixed_precision_training: Native AMP

Training results

Framework versions

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