Dialogue Mechanics: Tags, Beats, and Voice Gap

Dialogue Ratio tells you how much of a passage is speech. These three readings tell you what the speech is doing. They appear for fiction and screenwriting only, and all three stay hidden until dialogue makes up at least 10 percent of the passage, because measuring the texture of four lines of speech tells nobody anything.

Attribution Verb Variety

After every closing quote the engine looks ahead a short distance for an attribution verb, drawn from a list of 174. Your score is unique verbs divided by total attributions. It also reports how much of your attribution is said, says, asked, or asks, and how much comes from a list of 50 purple alternatives.

The benchmark runs 0.30 to 0.80 for both fiction and screenwriting, and the ceiling is the interesting end. Too much variety reads as attribution avoidance, where every line carries a different verb because the writer was afraid to repeat said. Said is close to invisible on the page. That is its job.

Two things that will skew your score

It is a ratio, so it moves against volume. Three tags using three different verbs scores a perfect 1.0. Forty tags using twelve verbs scores 0.30. A short scene will score higher than a long one written with identical habits, and neither number means what it appears to mean.

Only attribution that follows the quote is counted. If you write She said, “Hello,” with the tag in front, the engine does not see it. Writers who front-load tags will show fewer attributions than they actually wrote.

Dialogue-Action Beat Ratio

A block is a run of consecutive dialogue lines with no narration between them. The engine checks whether each block has an action beat immediately before or after it, detecting beats by looking for any of 818 action verbs in the narration gap. Your score is the share of blocks that sit next to a beat.

It also counts talking heads: blocks of four or more consecutive lines with no beat anywhere near them. That count is usually more useful than the ratio, because it points at specific stretches where the bodies have disappeared and two voices are exchanging lines in a white room.

The benchmark is a minimum of 0.50 for both types. At least half your dialogue blocks should touch a beat. Be aware that the stated maximum of 1.00 is also the highest value the metric can produce, so the warning about beats on every line cannot actually fire. If your score is 1.00, read the passage yourself and decide whether the pace has gone staccato.

Beat detection is word-list lookup, which cuts both ways. A beat written with a verb outside the list is invisible, and an action verb inside a tag (she turned to him and said) registers as a beat whether or not it does the work of one.

Dialogue Naturalness

The name oversells it, so here is what it actually measures: the gap between how your characters talk and how your narration talks. It compares three things across the two, average word length, average sentence length, and contraction rate, and averages the differences into a single number between 0 and 1.

It does not evaluate whether dialogue sounds like real speech. Nothing in the engine can do that. The benchmark runs 0.15 to 0.80. Below the floor, character voice has collapsed into narrating voice. Above the ceiling, characters read as broadly cartoonish, every line a costume.

When to disregard it

Three cases, and they are common enough to matter. All dialogue is pooled together, so the metric cannot distinguish one character from another; five distinct voices and five identical ones can score the same. Each signal is floored at zero, so narration that is more colloquial than the dialogue registers as no gap at all. And a first-person colloquial narrator scores low by construction, because the narration already talks the way the speech does. That is the book working, not failing.

What to do with the three

Read the talking heads count first, then the attribution breakdown, then the voice gap. That order runs from most concrete to most abstract, and the first two point at passages you can open and look at.

Open a scene in the editor and run Grade this passage with Fiction as the writing type. If dialogue is under 10 percent of the scene, these readings will not appear at all, which is the engine declining to measure something it cannot see enough of.