whisper-medium-et-ERR2020

This model is a fine-tuned version of openai/whisper-medium on the following training sets: Common Voice 11, VoxPopuli, FLEURS and ERR2020. The checkpoint-7000 was on Whisper Event leaderboard. Current score is for the final checkpoint.

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: 32
• eval_batch_size: 16
• 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: 10000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.1828	0.1	1000	0.3547	20.8829
0.09	0.2	2000	0.3476	19.0096
0.083	0.3	3000	0.3386	18.1304
0.0765	0.4	4000	0.3365	17.2591
0.0592	0.5	5000	0.3534	19.0213
0.0672	0.6	6000	0.3622	18.4263
0.0629	0.7	7000	0.3487	15.9839
0.0546	1.03	8000	0.3677	16.1021
0.0459	1.13	9000	0.3704	17.9073
0.0425	1.23	10000	0.3672	15.9119

The validation set is combined from the validation sets of Common Voice 11, VoxPopuli, FLEURS and ERR2020.

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.12.1+rocm5.1.1h
• Datasets 2.7.1.dev0
• Tokenizers 0.13.2