AhmedReda25300/SpeechX
Full fine-tune (fewer layers) vs openai/whisper-base. AhmedReda25300/SpeechX matches part of openai/whisper-base (97 of its 245 tensors; the rest of openai/whisper-base has no counterpart). On the shared part: Fine-tune: 96% of the weights changed, across 6 of 6 layers.
Tensors changed: other 4/4, Embeddings 0/1, Norms 26/26, MLP 24/24, Attention 42/42