Rakib/whisper-small-bn-crblp

Whisper Small - Mohammed Rakib

This model is a fine-tuned version of openai/whisper-small on the common-voice-11, the google-fleurs, the openslr53 and the crblp speech corpus datasets. It achieves the following results on the evaluation set:

• Loss: 0.0617
• Cer: 5.4436
• Wer: 9.6538

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: 4
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 16
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 8000
• training_steps: 40000

Training results

Training Loss	Epoch	Step	Validation Loss	Cer	Wer
0.5361	0.13	1000	0.4043	22.6599	44.0521
0.2881	0.26	2000	0.2217	16.3939	32.4894
0.2265	0.38	3000	0.1728	13.0425	25.9637
0.1974	0.51	4000	0.1430	11.3260	22.3187
0.1591	0.64	5000	0.1255	10.0167	19.5115
0.1504	0.77	6000	0.1102	8.8333	17.1919
0.1259	0.89	7000	0.1003	8.1863	15.8576
0.1184	1.02	8000	0.0940	7.7868	14.9110
0.1099	1.15	9000	0.0885	7.3675	13.9444
0.1075	1.28	10000	0.0830	6.9648	13.2008
0.095	1.41	11000	0.0789	6.6969	12.6776
0.0943	1.53	12000	0.0766	6.3765	11.9896
0.0923	1.66	13000	0.0731	6.1784	11.7203
0.0824	1.79	14000	0.0699	5.9267	11.1632
0.0756	1.92	15000	0.0683	5.6305	10.6327
0.0634	2.04	16000	0.0671	5.6905	10.6947
0.0618	2.17	17000	0.0662	5.5107	10.2926
0.0679	2.3	18000	0.0643	5.4948	10.1792
0.0589	2.43	19000	0.0647	5.5201	10.1881
0.0623	2.56	20000	0.0633	5.2731	9.8449
0.0558	2.68	21000	0.0623	5.4211	10.0267
0.0564	2.81	22000	0.0617	5.4553	9.9893
0.0552	2.94	23000	0.0607	5.3860	9.7778
0.0403	3.07	24000	0.0621	5.7297	10.0382
0.0406	3.19	25000	0.0617	5.4436	9.6538
0.041	3.32	26000	0.0611	6.0867	10.3834
0.0388	3.45	27000	0.0614	6.1641	10.3890
0.0383	3.58	28000	0.0611	6.1460	10.3537
0.0401	3.71	29000	0.0603	6.9576	11.0697
0.0343	3.83	30000	0.0613	7.1918	11.2243
0.0357	3.96	31000	0.0603	7.3128	11.3313
0.0313	4.09	32000	0.0624	7.3871	11.3861
0.0281	4.22	33000	0.0626	7.8705	11.8248
0.0298	4.34	34000	0.0629	8.3360	12.2368
0.0282	4.47	35000	0.0627	8.7840	12.6270

Framework versions

• Transformers 4.30.0.dev0
• Pytorch 2.0.0+cu117
• Datasets 2.10.2.dev0
• Tokenizers 0.13.2