sgangireddy/whisper-small-fi-full

Whisper Small Finnish all

This model is a fine-tuned version of openai/whisper-small on the mozilla-foundation/common_voice_11_0 fi dataset. It achieves the following results on the evaluation set:

• Loss: 0.5334
• Wer: 25.4333

Model description

The Model is fine-tuned for 5000 steps/updates on CV11 Finnish train+valiation data.

• Zero-shot - 30.5 (CV9 test data, even on CV11 the WER is closer a bit higher than this)
• Fine-tuned - 25.43 (CV11 test data)

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: 64
• 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
• training_steps: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0025	19.0	1000	0.4265	24.8493
0.0005	38.0	2000	0.4785	25.3203
0.0003	57.01	3000	0.5073	25.3956
0.0002	76.01	4000	0.5253	25.4333
0.0002	96.0	5000	0.5334	25.4333

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.0+cu117
• Datasets 2.7.1.dev0
• Tokenizers 0.13.2