arijitx/wav2vec2-xls-r-300m-bengali

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"
• —modelnameor_path="facebook/wav2vec2-xls-r-300m"
• —datasetconfigname="SLR53"
• —output_dir="./wav2vec2-xls-r-300m-bengali"
• —overwriteoutputdir
• —numtrainepochs="50"
• —perdevicetrainbatchsize="32"
• —perdeviceevalbatchsize="32"
• —gradientaccumulationsteps="1"
• —learning_rate="7.5e-5"
• —warmup_steps="2000"
• —lengthcolumnname="input_length"
• —evaluation_strategy="steps"
• —textcolumnname="sentence"
• —charstoignore , ? . ! \- \; \: \" “ % ‘ ” � — ’ … –
• —save_steps="2000"
• —eval_steps="3000"
• —logging_steps="100"
• —layerdrop="0.0"
• —activation_dropout="0.1"
• —savetotallimit="3"
• —freezefeatureencoder
• —featprojdropout="0.0"
• —masktimeprob="0.75"
• —masktimelength="10"
• —maskfeatureprob="0.25"
• —maskfeaturelength="64"
• —preprocessingnumworkers 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.