itcollectivelab/whisper-large-v3-ga-nss-augmented

Whisper Large v3 Performance Comparison

Standard Ghanaian vs Non-Standard Ghanaian Speech

Training Configuration	Validation WER	Validation CER	Test WER	Test CER
Standard Speech (SS)	0.1707	0.0935	0.1724	0.09
NSS - Full Tune	0.4487	0.2201	0.5450	0.3007
NSS - Frozen Encoder	0.5228	0.2479	0.56	0.3

Notes:

• Standard Speech: Trained on clean standard Ghanaian speech
• NSS - Full Tune: Both encoder and decoder trainable on non-standard speech
• NSS - Frozen Encoder: Encoder frozen, only decoder trainable on non-standard speech