ntviet/whisper-small-hre5.2

Whisper Small Hre 5.2, ASR for male & female Hre voice, 1000 steps, metric CER

This model is a fine-tuned version of openai/whisper-small on the Hre audio dataset 8 dataset. It achieves the following results on the evaluation set:

• Loss: 0.3295
• Cer Ortho: 19.5481
• Cer: 17.6265

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: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: constant_with_warmup
• lr_scheduler_warmup_steps: 50
• training_steps: 1000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Cer Ortho	Cer
3.6392	0.2304	50	3.0128	47.4354	44.3054
1.4265	0.4608	100	0.8819	47.9017	44.5998
0.5115	0.6912	150	0.6276	33.4828	30.3220
0.4144	0.9217	200	0.5343	28.6765	26.4029
0.2632	1.1521	250	0.4690	33.9491	31.7571
0.2403	1.3825	300	0.4225	23.8702	21.4535
0.2258	1.6129	350	0.4182	24.5158	22.3735
0.2089	1.8433	400	0.3895	23.4397	21.2144
0.1593	2.0737	450	0.3659	23.4935	21.2695
0.1206	2.3041	500	0.3742	27.4928	25.1150
0.1077	2.5346	550	0.3516	21.9512	19.9080
0.1122	2.7650	600	0.3224	29.5194	27.3597
0.1026	2.9954	650	0.3319	21.3056	18.8960
0.0524	3.2258	700	0.3371	21.3773	18.9144
0.0574	3.4562	750	0.3096	20.2116	18.2889
0.06	3.6866	800	0.3139	27.9232	26.0902
0.0538	3.9171	850	0.3194	20.0861	18.0497
0.0294	4.1475	900	0.3162	20.5524	18.4177
0.0281	4.3779	950	0.3343	20.4986	18.3809
0.0266	4.6083	1000	0.3295	19.5481	17.6265

Framework versions

• Transformers 4.51.3
• Pytorch 2.6.0+cu124
• Datasets 3.5.1
• Tokenizers 0.21.1