How to analyze a political speech: rhetoric, delivery, and measured signals
← All articles · By the RedFlag team · September 22, 2026
A political speech is the most watched and least carefully read piece of video in public life. Most commentary about one collapses two separate questions into a single verdict: was that true? and did he look like he believed it? The first is a matter of record. The second is a matter of delivery — and delivery, on a podium, is the hardest kind of video there is to read honestly, because everything about the format is designed to be controlled. This is a method for doing both jobs separately, and for knowing what the measured part can actually support.
RedFlag is not a lie detector. No tool, and no method on this page, can tell you whether a politician is telling the truth. What follows measures change in a speaker's own behavior over the course of a video. Whether a change means anything is your judgment, made against the record — never the tool's.
Layer 1: the argument, read on the transcript
Do this first, and do it without the video playing. Pull the transcript and mark, line by line, the structure underneath the language:
- The claim and its warrant. What is asserted, and what is offered as the reason to accept it? Political rhetoric very often supplies a vivid example where a warrant should be.
- The devices. Anaphora (repeated openings), tricolon (the three-part list), antithesis, the rhetorical question, the enumerated "plan." These are craft, not evidence of anything — a well-built sentence is equally available to a true statement and a false one.
- The hedges and the agentless passive. "Mistakes were made," "there were issues with the process," "to the extent that anything improper occurred." Note where the sentence stops naming who did what. That is a textual observation you can quote, and it is far stronger evidence than anything on the speaker's face.
- The checkable claims. List them. This is the part that gets resolved by sources, not by watching.
Only after the transcript is marked should you watch. Reading the rhetoric first keeps a confident delivery from lending credibility to a weak argument, which is exactly what a confident delivery is for.
Layer 2: the delivery — and why prepared remarks tell you almost nothing
Here is the part most guides get wrong. A prepared speech read from a teleprompter has been drafted by someone else, rehearsed, timed and marked up for emphasis. Its pace is even, its pitch contour is practiced, its pauses are deliberate. There is nothing to detect in it, because almost nothing in it is being produced in the moment.
That does not make the prepared section useless. It makes it the baseline. The useful evidence is what happens when the speaker leaves the script — the press Q&A, the hostile follow-up, the hot-mic aside, the debate rebuttal. A change from the speaker's own prepared-remarks baseline to their unscripted answers is a real, measurable observation. "He sounded flat" measured against nothing is not.
Check the shot before you trust the signals
Political video comes in specific formats, and the format determines which signals are even available. RedFlag makes this explicit rather than quietly guessing: the analysis runs when the face occupies at least 10% of frame height with a landmark-detector confidence of 0.35, but the blink channel only records data when the face fills at least 22% of the frame. Below that, the eyes are a handful of pixels and any blink rate you compute is noise — so the read rides on the voice channel alone, and the app says so rather than inventing a number. The thresholds are published in full on the methodology page.
| Format | What you get | What to do |
|---|---|---|
| Wide podium / parliament / chamber camera | Voice signals; face usually too small for blink data | Analyze it for pitch and voiced continuity; do not let anyone tell you what the blinking meant |
| Broadcast press conference, medium shot | Both channels, usually | The strongest political format to analyze — and Q&A is unscripted by definition |
| Sit-down interview or studio debate | Both channels, clean audio | Ideal; see how to analyze an interview |
| Rally footage with crowd noise | Degraded audio; pitch estimates suffer on overlapping sound | Weak material — prefer a clean feed of the same speech if one exists |
| Clipped 20-second excerpt | Not enough video to form a baseline | Don't. A baseline needs minutes, not seconds |
The method, step by step
- Get the full video, not the clip that is circulating. Selective editing is the single most common distortion in political footage, and a cut removes exactly the context a baseline is built from. If you only have the excerpt, say so out loud when you present anything you found.
- Mark the transcript as above, before watching.
- Identify the scripted/unscripted boundary. Usually the moment the speaker stops reading and takes the first question. Note the timestamp.
- Analyze the whole video in one pass. RedFlag computes each signal against that speaker's own median across the video, so the prepared portion becomes the reference the Q&A is measured against automatically — which is why you want both halves in one run rather than analyzing the Q&A on its own.
- Look only at flags in the unscripted portion. A flag during a teleprompter passage usually tells you about the writing (a long, flat, list-heavy paragraph) rather than about the speaker.
- Rewatch each flagged stretch with the transcript open. What was the question? What is the claim in that answer? Is it one of the checkable items on your list? This is where the two layers finally meet — and the only responsible way to present the result is "here is a specific moment, here is what the measurement was, here is what the record says," with all three stated separately.
What political delivery does to each signal
| Observation | The ordinary political explanation | What it is not |
|---|---|---|
| Flat, monotone stretch | Reading; a memorized talking point; an answer given five hundred times before | Evidence of a lie — rehearsal is the most common cause of monotony, and rehearsal is the job |
| Sudden pitch flattening in the Q&A | Careful wording under legal or political exposure; cognitive load while composing | A confession — careful wording is also what an honest person does when the stakes are high |
| Fewer pauses, continuous speech | Filibustering the clock, or refusing to give an interviewer an opening | Deception — it's a media tactic, and a well-known one |
| Blink elevation above their own rate | Stage lighting, dry air, fatigue at the end of a long day, concentration | Guilt — and on most podium shots it isn't even measurable |
| Filler words spike | An unanticipated question being answered in real time | Fabrication — composing a true answer produces the same fillers |
The research behind that middle column is not a matter of opinion: DePaulo et al.'s 2003 review of over a hundred cue studies found no single behavioral marker that reliably indicates deception, and Bond & DePaulo's 2006 meta-analysis put human judgment of truthfulness at about 54% accuracy — a result you can test on yourself in a few minutes with the 54% test. What behavior does track, per Vrij, Fisher & Blank (2017), is cognitive load: effortful recall, careful phrasing and self-monitoring compete for the same mental resources whether the underlying statement is true or false.
A worked example
For a full analysis run on a public denial where the documented record was later settled, see our walkthrough of Lance Armstrong's 1999 doping denial: 6 flagged moments, driven almost entirely by pitch flattening rather than anything facial. It is a useful calibration for political analysis precisely because it is modest — the signals marked where to look, and the historical record, not the tool, supplied the conclusion.
Publishing what you find, responsibly
If the speech you are analyzing involves a living person and unresolved allegations, the honest form is narrow: name the moment, state the measurement, state what is and is not established. Do not write that the measurement means the person lied; it cannot mean that, and claiming otherwise is both wrong and defamatory in a way the underlying data does not support. The interesting finding in this kind of work is almost always where the delivery changed, and what question caused it — which is journalism, not mind-reading.
Analyze a speech yourself. Paste a public YouTube URL into the RedFlag analyzer — free with no signup (1 video a day, up to 10 minutes; 2 a day and 30 minutes once you sign in), and the analysis runs in your browser. RedFlag is not a lie detector: it shows you where a speaker's signals diverged from their own baseline, and leaves the judgment to you.
Open the analyzerHow the scoring works · How to analyze an interview · Test your own judgment
Sources
- Bond, C. F., & DePaulo, B. M. (2006). Accuracy of deception judgments. Personality and Social Psychology Review, 10(3), 214–234.
- DePaulo, B. M., et al. (2003). Cues to deception. Psychological Bulletin, 129(1), 74–118.
- Vrij, A., Fisher, R. P., & Blank, H. (2017). A cognitive approach to lie detection. Legal and Criminological Psychology, 22(1), 1–21.
- RedFlag scoring method, face-size gates and thresholds: How RedFlag scores a video.