Akashpb13/Swahili_xlsr

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - hu dataset. It achieves the following results on the evaluation set (which is 10 percent of train data set merged with dev datasets):

Loss: 0.159032
Wer: 0.187934
Model description
"facebook/wav2vec2-xls-r-300m" was finetuned.

Intended uses & limitations

More information needed

Training and evaluation data

Training data - Common voice Hausa train.tsv and dev.tsv Only those points were considered where upvotes were greater than downvotes and duplicates were removed after concatenation of all the datasets given in common voice 7.0

Training procedure

For creating the training dataset, all possible datasets were appended and 90-10 split was used.

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.000096
train_batch_size: 16
eval_batch_size: 16
seed: 13
gradient_accumulation_steps: 2
lr_scheduler_type: cosine_with_restarts
lr_scheduler_warmup_steps: 500
num_epochs: 80
mixed_precision_training: Native AMP

Training results

Step	Training Loss	Validation Loss	Wer
500	4.810000	2.168847	0.995747
1000	0.564200	0.209411	0.303485
1500	0.217700	0.153959	0.239534
2000	0.150700	0.139901	0.216327
2500	0.119400	0.137543	0.208828
3000	0.099500	0.140921	0.203045
3500	0.087100	0.138835	0.199649
4000	0.074600	0.141297	0.195844
4500	0.066600	0.148560	0.194127
5000	0.060400	0.151214	0.194388
5500	0.054400	0.156072	0.192187
6000	0.051100	0.154726	0.190322
6500	0.048200	0.159847	0.189538
7000	0.046400	0.158727	0.188307
7500	0.046500	0.159032	0.187934

Framework versions

Transformers 4.16.0.dev0
Pytorch 1.10.0+cu102
Datasets 1.18.3
Tokenizers 0.10.3

Evaluation Commands

To evaluate on mozilla-foundation/common_voice_8_0 with split test

python eval.py --model_id Akashpb13/Swahili_xlsr --dataset mozilla-foundation/common_voice_8_0 --config sw --split test

Akashpb13
/

Swahili_xlsr