Modelpublic

RuudVelo/wav2vec2-large-xls-r-1b-cv8-mt

sourceHugging Faceapache-2.0updated 5y agoView on Hugging Face

0likes82downloads

Model Card

wav2vec2-large-xls-r-1b-cv8-mt

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the common_voice dataset. It achieves the following results on the evaluation set:

—Loss: 0.2210
—Wer: 0.1974

Model description

Note: another version of this model is available with a KenLM 3gram model. This model performs better than this model. See https://huggingface.co/RuudVelo/wav2vec2-large-xls-r-1b-cv8-mt-lm

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following config and hyperparameters were used during training:

model = Wav2Vec2ForCTC.frompretrained( "facebook/wav2vec2-xls-r-1b", attentiondropout=0.05, hiddendropout=0.05, featprojdropout=0.05, masktimeprob=0.55, maskfeatureprob=0.10, layerdrop=0.05, ctczeroinfinity=True, ctclossreduction="mean", padtokenid=processor.tokenizer.padtokenid, vocabsize=len(processor.tokenizer), )

from transformers import TrainingArguments

trainingargs = TrainingArguments( outputdir=reponame, groupbylength=True, perdevicetrainbatchsize=32, gradientaccumulationsteps=2, evaluationstrategy="steps", numtrainepochs=50, gradientcheckpointing=True, fp16=True, savesteps=400, evalsteps=400, loggingsteps=400, learningrate=5.5e-05, warmupsteps=500, savetotallimit=2, pushtohub=True, report_to="tensorboard")

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.4564	13.33	400	0.3783	0.3981
0.7931	26.66	800	0.2377	0.2298
0.5364	39.98	1200	0.2210	0.1974

Note that the test WER of 19.74 is different than the above reported 17.57. This was due to a bug which was found while processing files with an older version of the datasets library. The right library is listed below.

Framework versions

—Transformers 4.17.0.dev0
—Pytorch 1.10.2+cu102
—Datasets 1.18.3
—Tokenizers 0.11.0