Your AI Voice Says the Number Before Anyone Has Found the Axis: How to Narrate Charts and Data Videos (2026)

1. The two-stage chart read — and the timing mismatch

When a chart appears on screen, a viewer does two quite different things in sequence.

Stage one — shape, roughly 1–2 seconds, effortless. They see that something goes up, or that one bar is much taller than the rest, or that the dots form a cloud with a tilt. This is pre-attentive: it happens whether they want it to or not, and it costs almost nothing.

Stage two — detail, roughly 5–10 seconds, effortful. They read the axis labels, work out the units, find the legend, locate the specific point you are talking about, and check the scale. This is slow, deliberate, and it is where a viewer can be lost.

Now look at what a typical generated data script does. Handed a dataset and asked for narration, a model produces something like:

"Global electric vehicle sales reached fourteen point two million units in 2024, up from thirteen point nine million in 2023, representing a two point two percent increase."

That is accurate, fluent and delivered at conversational pace, which means the entire sentence is over in about eight seconds. In those eight seconds the viewer has completed stage one and is somewhere in the middle of stage two — still hunting for the axis units. They heard "fourteen point two million" while their eye was on the legend. The number arrived before the container it was supposed to go into.

And because nothing sounded wrong, the viewer does not attribute the failure to the script. They attribute it to themselves — "I'm not good with data" — and they leave.

Three consequences follow, and the rest of this article is built on them:

• There must be a gap between a chart appearing and the first word about its contents.
• The voice must state the shape before it states any number, because the shape is what the viewer already has.
• A single chart can carry far fewer spoken numbers than a script writer instinctively wants to put on it.

2. The orientation beat

The orientation beat is a deliberate pause between the chart appearing and the narration resuming. It is the single highest-return change most data channels can make, and it costs nothing but seconds.

What goes in the beat? Ideally nothing but music or room tone. If total silence over a static chart makes you uncomfortable, the acceptable filler is a framing phrase that contains no data — "This is where it gets interesting", "Watch the right-hand side" — spoken before the chart appears, so the beat itself stays clear.

What must never go in the beat: the first number, the axis units, or the punchline. Those all arrive after it.

The exception: animated chart builds

If the chart draws in — bars growing, a line sweeping left to right — the orientation beat overlaps the animation. The rule becomes: say nothing until the animation completes, plus 0.5 seconds. Narrating over a building chart is the same error in a different costume, because the viewer's stage-one read cannot complete until the shape has stopped changing.

3. Shape first, number second

After the orientation beat, the first thing out of the voice should describe what the viewer has already seen — the shape. This does two jobs: it confirms they read it correctly, and it tells them which part of the chart the numbers are about to refer to.

Then the number. Then, only if needed, the interpretation.

Notice the last row. The generated version reads out four numbers; the rewrite reads out zero and says more. That is not a stylistic preference — it is the point of the next section.

4. One number per chart (the comparison is the message)

Here is the rule that most improves data narration and that almost nobody follows: speak at most one number per chart, sometimes two, and let the chart carry the rest.

The reasoning is simple. A spoken number is expensive — it occupies attention, it cannot be re-read, and it has to be held in working memory to be useful. A number on screen is cheap: it sits there, the viewer can glance back, and it costs nothing to display fifty of them. Putting a number in the voice is therefore a decision to spend something scarce, and you should spend it only on the number the argument turns on.

What the voice should carry instead of the numbers:

• The comparison. "Twice the next one." "Roughly the same as everything else combined." "Back where it was in 2016."
• The direction and whether it is surprising. "Up, which everyone expected." "Down, which nobody did."
• The scale translation. "Four hundred thousand people — about the population of Minneapolis."
• The caveat. "This axis doesn't start at zero." "These are estimates after 2023."
• Where to look next. "Now the right-hand end."

When you genuinely do need to walk through several values — a ranking, a step-by-step build — treat it as a list with pauses rather than a sentence: name, number, 400–600 ms, name, number, 400–600 ms. A comma-separated run of five figures spoken at conversational pace is retained by nobody.

5. Number-reading conventions your engine gets wrong

Text-to-speech models were trained overwhelmingly on prose, and data notation is not prose. The failures below are common, and nearly all of them are silent — the engine does not error, it produces a fluent sentence that says the wrong thing.

The three that cause actual errors, not just awkwardness

1. Percent versus percentage points. "Unemployment rose four percent" and "unemployment rose four percentage points" are wildly different claims, and a dropped "pp" silently converts one into the other. Never let the abbreviation into the audio script.
2. Dropped negation and dropped comparators. "Not statistically significant" losing its "not", or "≥" vanishing, changes the meaning completely while leaving a fluent sentence behind. Search for these specifically.
3. Dropped uncertainty markers. "±", "~", "est.", "projected" are all frequently swallowed. The result is a video that states forecasts as measurements — which is a credibility problem, not a pronunciation problem.

6. The chart-type playbook

Different charts hand the viewer different things for free, so the voice has a different job in each.

Two rows deserve underlining. Log axes need re-stating every time they appear, not once at the start — viewers join late, viewers forget, and a log axis narrated as if it were linear is the most consequential honest-mistake in data video. And bar chart races should be nearly silent; they are already doing the thing your voice would be doing, and talking over them is the clearest sign a script was written without watching the footage.

7. Speech density for data segments

Density here means the share of a segment's running time that contains speech. Budget it before writing.

Rate matters less than you would think — 145–155 wpm is fine for data narration, and slowing to 120 does not fix comprehension. The variable that fixes comprehension is when you speak, not how fast.

There is one rate exception worth building in: drop 10–15% for the sentence containing the single spoken number. Not for drama — for parsing. "Fourteen point two million" needs more time per syllable than the words around it, and a uniform rate is what makes multi-part numbers arrive as a blur.

Why this is newly practical

Writing narration to measured gaps used to be a luxury: if a line ran 1.3 seconds long you either re-recorded it or you lived with it. Re-rendering the same sentence at 148, 140 and 134 words per minute to fit a 9.2-second static chart costs nothing with text-to-speech. So the correct working order for data video inverts the usual one: build the charts, measure their on-screen times, then write to those durations.

8. Honesty rules: what the voice must not claim

Data narration has a specific integrity risk that other narration does not: a confident voice can assert a relationship the chart does not contain, and viewers will believe the voice over the picture. A few hard rules.

• Do not narrate causation over a correlation chart. "Which is why" and "because of this" are causal claims. If the chart shows two variables moving together, the voice says they move together.
• Say when the axis is truncated. Every time it happens. "Note the axis starts at sixty, not zero" costs two seconds and is the difference between explaining and misleading.
• Say when data is projected, estimated or modelled. Especially where the chart switches from solid to dashed line — narrate the switch: "everything from here is a projection".
• Do not read a precision you do not have. If the underlying figure is a survey estimate with a wide interval, "forty-one point three percent" implies a precision that does not exist. Say "about forty percent".
• Attribute the source aloud once per dataset. "This is from the [organisation]'s 2025 release." On screen is good; in the voice is better, because the voice is what gets quoted back.
• Do not let a rewrite change a number. When you compress a sentence for timing, the number is the one element that may not be adjusted for rhythm. Re-check every figure against the source after the timing pass, not before.

General information only. Nothing here is financial, medical or investment advice, and data videos covering markets, health or public policy should carry their own appropriate disclaimers.

9. Script prep: the sweep that catches silent failures

Run these passes on every finished script, in this order, before rendering anything.

1. Digit sweep. Search for every numeral 0–9. Each hit is either converted to words or deliberately left. This catches the majority of problems on its own.
2. Symbol sweep. Search for: % pp ± ~ × ÷ ≥ ≤ ≠ $ £ € ¥ ² ³ ° / – — → ° and the letters k, M, B, bn, tn when attached to a number. Convert each.
3. Negation and comparator sweep. Search "not", "no", "never", "without", "less than", "fewer", "more than", "at least". Confirm each survives in the rendered audio — these are the highest-consequence silent drops.
4. Uncertainty sweep. Search "est", "approx", "projected", "forecast", "modelled", "±", "~". Confirm each is spoken.
5. Acronym ruling. Build a list of every acronym in the script and decide, once, whether each is spelled ("C A G R"), said as a word ("NATO"), or expanded. Write the ruling into a channel lexicon so it is consistent across videos.
6. Unit lexicon. Same treatment for units — km², kWh, bbl/d, m³, per capita, per hundred thousand. Add them to the lexicon file.
7. One-number check. Go chart by chart and count the spoken numbers. Any chart with more than two, justify or cut.
8. Beat insertion. Add the orientation beat before every chart's first line, per the table in section 2.

10. Workflow and the double QC test

1. Build or export the charts first.
2. Cut them to length with no narration. Write down each chart's on-screen duration.
3. Assign a density and calculate a word budget: seconds × density × 2.4 words.
4. Write to the budget, shape-first, one number per chart.
5. Run the eight sweeps in section 9.
6. Render segment by segment, not as one block, so any line can be re-timed independently.
7. Lay in with the orientation beats. Re-render overruns at a higher rate rather than trimming in the edit.
8. Run the double test below.

The double test

Picture off. Listen to the narration alone. It should still be a coherent argument — but it should contain obvious holes where a chart is doing work ("the one that fell is the one that used to be biggest" is meaningless without the picture, and that is correct). If the audio alone is a complete, self-sufficient explanation, you are over-narrating and the charts are decoration.

Sound off. Watch the video muted. If you can extract the argument entirely, the voice is redundant. If you can extract nothing, the charts are under-labelled and the voice is compensating for a design problem.

You want both tests to fail in the right direction: each channel incomplete alone, complete together. That is what "the chart and the voice are doing different jobs" actually means in practice.

And one transcription pass

Run the rendered audio back through speech-to-text and diff it against your script. Number errors, dropped negations and swallowed symbols are frequently inaudible to someone who knows what the line was supposed to say — you hear the sentence you wrote, not the sentence that was rendered. A diff does not have that problem. Missing "not", a changed digit and a vanished "percentage points" are the three hits that matter.

11. FAQ

Because "too fast" almost always means "the number arrived before I had found it on the chart". The fix is timing, not rate: insert an orientation beat of 1.5–3.5 seconds between a chart appearing and the first word about it, and state the shape before any figure. Slowing the voice to 120 words per minute without fixing the beat changes nothing.

One, sometimes two. Spoken numbers occupy scarce attention and cannot be re-read; on-screen numbers are free and persistent. Spend the spoken slot on the figure the argument turns on, and let the voice carry the comparison, the direction, the scale translation and the caveat instead.

A rise from 4% to 6% is two percentage points and a 50% increase — very different claims. The abbreviation "pp" is routinely mangled or dropped by text-to-speech engines, which silently converts one claim into the other while leaving a fluent sentence. Always write the full words "percentage points" into the audio script, and never let the abbreviation through.

Do not rely on the engine. Run a symbol sweep across the finished script — search for % pp ± ~ × ≥ ≤ ≠ ² ³ and currency signs — and convert every instance to words before rendering. These failures are silent: the symbol is simply dropped and the sentence still sounds correct, which is why they survive listening checks and only show up in a transcribe-and-diff pass.

No. A viewer's shape-reading cannot complete while the shape is still changing, so narration during a build is wasted. Set up before it starts, stay silent through it plus about half a second, then land the point. Bar chart races in particular should run almost entirely unnarrated — they are already performing the function your voice would be performing.

145–155 words per minute is fine for the general narration. Drop 10–15% for the single sentence containing your spoken number, because multi-part figures need more time per syllable than surrounding prose. Beyond that, when you speak matters far more than how fast — an orientation beat buys more comprehension than any rate change.