M. Zusag

Achieving a 79% F1 score in disfluency detection with zero-shot learning demonstrates that ASR models can effectively control transcription style without extensive retraining.

Achieving a 98.7% reduction in speaker identification accuracy reveals how focusing on linguistic content can drastically enhance speech tokenization efficiency.