whisper-ro / README.md
udan11's picture
initial commit
a2fd10c unverified
metadata
license: apache-2.0
tags:
  - automatic-speech-recognition
  - whisper
  - romanian
datasets:
  - readerbench/echo
metrics:
  - wer
model-index:
  - name: whisper-ro
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Echo
          type: readerbench/echo
          config: ro
        metrics:
          - name: WER
            type: wer
            value: 0.08668345828147764

whisper-ro

This model is a fine-tuned version of openai/whisper-small on the Echo dataset, a large open-source Romanian dataset.

Name Small Large-v2 Fine-tuned small
(this model)
Common Voice 33.2 15.8 12.2
FLEURS 29.8 14.4 10.9
VoxPopuli 28.6 14.4 9.4
Echo >100 >100 8.6
RSC 38.6 28.5 5.4

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 1e-05
  • train_batch_size: 128
  • eval_batch_size: 128
  • seed: 42
  • distributed_type: multi-GPU
  • num_devices: 2
  • total_train_batch_size: 256
  • total_eval_batch_size: 256
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 500
  • num_epochs: 20.0
  • mixed_precision_training: Native AMP