Figure 6.
The horizontal axis shows Training Steps in thousands, from 0 to 90. The vertical axis shows alpha Value from 0.75 to 1.00. All four series begin at about 1.00 at 0 training steps and drop sharply by about 9 thousand steps. With A P E, alpha falls to about 0.78, rises to about 0.85 near 27 thousand steps, then gradually declines to about 0.80 at 90 thousand steps. With P prime subscript U T T T Q L, alpha falls to about 0.80, rises to a peak near 0.88 around 45 thousand steps, then declines gradually to about 0.85 at 90 thousand steps. With P prime subscript C H A R T Q L, alpha falls to about 0.80, rises to about 0.86 around 36 thousand steps, then decreases to about 0.83 at 90 thousand steps. With P prime subscript U T T plus C H A R T Q L, alpha falls to about 0.79, rises to about 0.86 around 36 thousand steps, then declines to about 0.83 at 90 thousand steps.Evolution of the learned scaling coefficient during training. TQL configurations converge to higher (–) than APE only (), indicating that explicit quality supervision enables more effective utilization of positional cues. Among TQL configurations, character-level TQL converges to lower than utterance-level TQL, as per character labels partially fulfill the role of positional encoding