whisper-twi-commands-small

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

Model description

A fine-tuned version of the OpenAI whisper small model.

For transcribing akan audio to akan text when using the sw tag. Limited to Asante and Akuapem dialects. Model struggles with words outside the Bible.

Training Loss	Epoch	Step	Validation Loss
3.8869	0.04	200	1.8523
1.6705	0.08	400	0.8196
1.2752	0.12	600	0.5912
0.9939	0.16	800	0.4853
0.8389	0.2	1000	0.4296
0.6369	1.0266	1200	0.3904
0.6299	1.04	1400	0.3836
0.6273	1.08	1600	0.3584
0.6677	1.12	1800	0.3435
0.6077	1.16	2000	0.3254
0.5247	1.2	2200	0.3140
0.3660	1.04	2400	0.3260
0.3733	1.08	2600	0.3161
0.2308	1.04	2800	0.3343
0.2429	1.08	3000	0.3296
0.1571	1.04	3200	0.3623
0.1557	1.08	3400	0.3513
0.4454	1.12	3600	0.3076
0.3961	1.16	3800	0.2999
0.3381	1.2	4000	0.2956
0.2404	2.0266	4200	0.3045
0.2702	2.0666	4400	0.3029
0.1122	1.04	4600	0.3314
0.0967	1.08	4800	0.3302
0.3191	1.12	5000	0.3140