This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the OPENSLR_SLR53 - bengali dataset. It achieves the following results on the evaluation set.

Without language model :

• WER: 0.21726385291857586
• CER: 0.04725010353701041

With 5 gram language model trained on 30M sentences randomly chosen from AI4Bharat IndicCorp dataset :

• WER: 0.15322879016421437
• CER: 0.03413696666806267

Note : 5% of a total 10935 samples have been used for evaluation. Evaluation set has 10935 examples which was not part of training training was done on first 95% and eval was done on last 5%. Training was stopped after 180k steps. Output predictions are available under files section.

Training hyperparameters

The following hyperparameters were used during training:

• dataset_name="openslr"
• model_name_or_path="facebook/wav2vec2-xls-r-300m"
• dataset_config_name="SLR53"
• output_dir="./wav2vec2-xls-r-300m-bengali"
• overwrite_output_dir
• num_train_epochs="50"
• per_device_train_batch_size="32"
• per_device_eval_batch_size="32"
• gradient_accumulation_steps="1"
• learning_rate="7.5e-5"
• warmup_steps="2000"
• length_column_name="input_length"
• evaluation_strategy="steps"
• text_column_name="sentence"
• chars_to_ignore , ? . ! - ; : " “ % ‘ ” � — ’ … –
• save_steps="2000"
• eval_steps="3000"
• 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"
• preprocessing_num_workers 32

Framework versions

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

Notes

• Training and eval code modified from : https://github.com/huggingface/transformers/tree/master/examples/research_projects/robust-speech-event.
• Bengali speech data was not available from common voice or librispeech multilingual datasets, so OpenSLR53 has been used.
• Minimum audio duration of 0.5s has been used to filter the training data which excluded may be 10-20 samples.
• OpenSLR53 transcripts are not part of LM training and LM used to evaluate.