Token Level Timing And Alignment Extraction

via “timestamp-aligned transcription with segment-level timing information”

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

Unique: Extracts timing from decoder attention weights without separate forced-alignment model — the cross-attention mechanism naturally learns to align generated tokens to input time-steps, enabling end-to-end timing in single pass rather than requiring post-hoc alignment

vs others: More efficient than two-pass approaches (transcribe then align) and eliminates dependency on separate alignment models like Montreal Forced Aligner; timing emerges naturally from the attention mechanism rather than being bolted on as post-processing