README.md

metadata

language:
  - ml
license: apache-2.0
tags:
  - whisper-event
  - hf-asr-leaderboard
  - generated_from_trainer
datasets:
  - mozilla-foundation/common_voice_11_0
metrics:
  - wer
model-index:
  - name: whisper_malayalam_largev2
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: mozilla-foundation/common_voice_11_0
          type: mozilla-foundation/common_voice_11_0
          config: ml
          split: test
        metrics:
          - name: Wer
            type: wer
            value: 68.7356

Whisper-fineTuning-malayalam

This model was trained from scratch on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.4138
• Wer: 68.7356

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: 1e-05
• train_batch_size: 12
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 4000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0099	1.88	1000	0.3563	69.6552
0.0046	3.77	2000	0.3860	70.1149
0.001	5.65	3000	0.4105	70.3448
0.0001	7.53	4000	0.4138	68.7356

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.0+cu116
• Datasets 2.8.0
• Tokenizers 0.13.2