whisper-medium-jp (ONNX)

This is an ONNX version of drepic/whisper-medium-jp. It was automatically converted and uploaded using this Hugging Face Space.

The optimal Q4 quantitative configuration: using decoder_model with q4f16 results in less ambiguous or nonsensical outputs.

quantization: {
    encoder_model: 'q4f16',
    decoder_model_merged: 'q4',
},

Usage with Transformers.js

This model is a fine-tuned version of openai/whisper-medium on an unknown dataset. It achieves the following results on the evaluation set:

More information needed

More information needed

More information needed

The following hyperparameters were used during training:

learning_rate: 4e-06
train_batch_size: 4
eval_batch_size: 2
seed: 42
distributed_type: multi-GPU
num_devices: 2
gradient_accumulation_steps: 2
total_train_batch_size: 16
total_eval_batch_size: 4
optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 400
num_epochs: 15
mixed_precision_training: Native AMP

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.5341	1.0	7155	0.5321	0.2416	0.2416
0.5023	2.0	14310	0.5143	0.2369	0.2369
0.499	3.0	21465	0.5063	0.2337	0.2337
0.4773	4.0	28620	0.5010	0.2310	0.2310
0.4775	5.0	35775	0.4944	0.2289	0.2289
0.4709	6.0	42930	0.4886	0.2288	0.2288
0.4907	7.0	50085	0.4870	0.2271	0.2271
0.4855	8.0	57240	0.4868	0.2261	0.2261
0.4487	9.0	64395	0.4828	0.2254	0.2254