infinitejoy's picture
robust speech event tag
68ff46b
|
raw
history blame
1.98 kB
metadata
language:
  - hy-AM
license: apache-2.0
tags:
  - automatic-speech-recognition
  - mozilla-foundation/common_voice_7_0
  - generated_from_trainer
  - robust-speech-event
datasets:
  - common_voice
model-index:
  - name: wav2vec2-large-xls-r-300m-armenian
    results: []

wav2vec2-large-xls-r-300m-armenian

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - HY-AM dataset. It achieves the following results on the evaluation set:

  • Loss: 0.9669
  • Wer: 0.6942

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 0.0003
  • train_batch_size: 32
  • eval_batch_size: 32
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 500
  • num_epochs: 200.0
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
1.7294 27.78 500 0.8540 0.9944
0.8863 55.56 1000 0.7282 0.7312
0.5789 83.33 1500 0.8178 0.8102
0.3899 111.11 2000 0.8034 0.7701
0.2869 138.89 2500 0.9061 0.6999
0.1934 166.67 3000 0.9400 0.7105
0.1551 194.44 3500 0.9667 0.6955

Framework versions

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