asr-africa/whisper_BIG-C_BEMBA_189hr_v1

Whisper Small Bemba - Beijuka Bruno

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

• Loss: 0.5013
• Wer: 0.3531
• Cer: 0.1008

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: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.025
• num_epochs: 100

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.9062	1.0	5914	0.4964	0.4258	0.1059
0.5025	2.0	11828	0.4104	0.3567	0.0887
0.4079	3.0	17742	0.3767	0.3252	0.0827
0.3239	4.0	23656	0.3676	0.3133	0.0804
0.2438	5.0	29570	0.3798	0.3219	0.0846
0.1655	6.0	35484	0.4092	0.3124	0.0787
0.0986	7.0	41398	0.4579	0.3251	0.0845
0.0554	8.0	47312	0.4980	0.3231	0.0844
0.0342	9.0	53226	0.5362	0.3174	0.0820
0.0255	10.0	59140	0.5647	0.3150	0.0810
0.021	11.0	65054	0.5882	0.3153	0.0797
0.0184	12.0	70968	0.6067	0.3162	0.0805
0.0161	13.0	76882	0.6337	0.3192	0.0842
0.0146	14.0	82796	0.6493	0.3138	0.0819

Framework versions

• Transformers 4.45.2
• Pytorch 2.1.0+cu118
• Datasets 3.0.1
• Tokenizers 0.20.1