Efficient Inference Through Optimized Transformer Architecture

via “efficient inference through knowledge distillation and model compression”

automatic-speech-recognition model by undefined. 67,92,170 downloads.

Unique: Uses knowledge distillation from full v3 model to compress parameter count by ~50% while preserving 99-language coverage through shared multilingual embeddings — the student model learns to match the teacher's output distributions rather than training from scratch, enabling faster convergence and better generalization

vs others: Faster than full Whisper v3 (2-3x speedup) while maintaining multilingual capability; more accurate than naive pruning approaches because distillation preserves learned representations; enables deployment scenarios (mobile, edge, real-time) where full model is infeasible