Skip to content

This repo contains the baseline model recipes and pre-trained model for GramVanni hindi ASR challenge

Notifications You must be signed in to change notification settings

gpu-poor/gramvaani_hindi_asr

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

33 Commits
 
 
 
 
 
 
 
 

Repository files navigation

gramvaani_hindi_asr

Baseline model's performance on Gram vaani 5h dev set

Model Type Framework Hyper-params Test Set WER CER LM
Transformer ESPnet 12 layers; nbpe: 1000 dev_5h 37.3 19.6 Not used
Conformer ESPnet 12 layers ; kernel size 15 ; nbpe: 1000 dev_5h 33.4 17.5 Not used
TDNN Kaldi 13 layers dev_5h 30.12 16.12 3-gram LM
CNN-TDNN Kaldi 12 layers tdnn with cnn front-end dev_5h 30.93 17.03 3-gram LM

Baseline Model Pretraining data Finetuning data Model dev WER
I Wav2Vec 2.0 Base 1000 hours Gramvaani unlabelled No finetuning download N.A
II Wav2Vec 2.0 Base 1000 hours Gramvaani unlabelled 100 hours Gramvaani labelled download 35.973
III Wav2Vec 2.0 Large 1000 hours Gramvaani unlabelled No finetuning download N.A
IV Wav2Vec 2.0 Large 1000 hours Gramvaani unlabelled 100 hours Gramvaani labelled download 35.870
V Wav2Vec 2.0 Base N.A (pretrained model from AI4Bharat) 100 hours Gramvaani labelled download 33.307
VI Wav2Vec 2.0 Base N.A (pretrained model from Open-Speech-EkStep) 100 hours Gramvaani labelled download 34.328

Clone this repository to try recipes and trained baseline models in various framework

Espnet based Transformer/Conformer recipe and trained baseline models are avilable in ./espnet/asr/

Kaldi based recipes are available in ./kaldi/asr

Wav2vec2.0 based models and config files are available at ./wav2vec2.0

About

This repo contains the baseline model recipes and pre-trained model for GramVanni hindi ASR challenge

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published