Call out the claims that don't hold up.
Live speech-to-text, real-time fact checks, and authentic competition formats — from casual to Mock Trial.
Debate Mode
Topic
Rounds & Timer
Debaters
Point System
AI Judging & Auto Fact-Check
—
Debater A
Debater B
Live Subtitles
Fact-Check
Highlights
Debate someone on another device.
Two accounts, two devices, one shared round. There's no shared audio channel between devices yet — for crossfire or discussion phases, coordinate by call or in person; Verdict Debate AI syncs the turns, transcript, and scores between you.
Score a recorded debate.
Drop in a video or audio file of a debate and we'll transcribe it, attribute each turn to a side, fact-check every claim, flag fallacies, and produce the same per-side breakdown the live mode gives you.
1. Pick a file
2. Tell us about the debate
3. Run the analysis
Transcribing takes roughly 60–90 seconds per 30 minutes of audio. Scoring is paced to stay within free-tier API rate limits — roughly 20 seconds per speaker turn, so a debate with many turns can take several minutes to fully score.
Fallacy library & format vocabulary.
Every fallacy flagged during a debate deep-links here. Look one up mid-round, or browse ahead of time.
Fallacy Library
Format Vocabulary
Build a case outline from scratch.
Answer a few questions and get a structured contention outline — taglines, warrants, and what evidence to go find — to build from, not to read verbatim.
Stress-test your case before the round.
Paste a case, a constructive speech, or an outline. The AI flags logical gaps, weak or unsupported warrants, and claims that need evidence — before a judge finds them for you.
1. Paste your case
How you're trending.
Every debate you finish (live or practice) gets logged locally in this browser — no account needed. See which fallacies you keep running into and how different judge personas have scored you.
See how your whole league is doing.
For coaches and tournament directors. Enter a league code to see aggregate stats across everyone submitting under it. Needs a Postgres database attached to this deployment — see the README.
From "they're speaking" to "point scored" — in under a second.
Say something true and the screen flashes green before you've finished your next breath. Say something false and it flashes red. Nobody's raising a hand to fact-check you — it's already done. Here's the machinery behind that, round by round.
1 Setup
Six formats, one screen. Pick Public Forum and the whole speech order snaps into place — constructives, crossfires, summaries, final focus, real time limits, zero guessing. Parliamentary runs the actual PMC → LOC → MG → MO → LOR → PMR sequence. Casual mode throws all of that out — you just set the rounds, the clock, and whether sides trade off.
Pick a topic — a preset or your own — name the debaters, set your spoken language, decide what points are worth, and choose whether an AI judge weighs in at the end.
2 Live speech capture
Hit start and the mic goes live. Words show up almost as fast as you say them — faded at first, while the recognizer's still deciding what it heard, then locked in solid once it's sure. That mic light isn't decoration; it's the only proof anyone's actually listening.
And while you're mid-sentence on your next point, the one you just finished is already getting checked behind the scenes.
3 Claim detection & fact-check
Every statement is sorted before anything happens:
- What kind of statement is it? A fact or statistic, an opinion, a value judgment, a prediction, a what-if, a claim about the case (Mock Trial), or something unverifiable. Views are logged with a label but never scored — you won't get dinged for saying "I think."
- Can it be settled right now? Well-known facts and statistics are judged on the spot. Harder ones — exact figures, recent events, niche sources — are held back and checked with a Wikipedia lookup at turn changes and after the round.
The checks are paced on purpose — sentences are grouped into one check every few seconds (longer when the AI service is busy), nothing already said is checked twice, and they go quiet during crossfire, because even we can't tell who's talking over whom.
4 Instant feedback
Facts that can be settled on the spot come back fast. Every statement gets a label:
- True or mostly true — screen flashes green, a short high tone plays, and the speaker gets
+fact(partial points for mostly true). - False or misleading — screen flashes red, a low tone plays, and the speaker loses
−wrong(partial for misleading). - Opinion, value claim, prediction, hypothetical, case claim, unverifiable or inconclusive — logged in the claim log with the reason, and no points move.
- To verify — held back until a pause or the end of the round. If it turns out true or false, points are added or removed then and the results page marks it "checked later."
A banner surfaces the claim in plain English, one line of reasoning, and whoever
just got scored. Every one of these calls is tagged · auto in the undo
tray, in case the machine got it wrong and you need to say so.
5 Manual scoring & fallacies
The AI doesn't get the final word — you do. Missed a claim? Hit +fact or −wrong yourself. Heard a straw man or a slippery slope slide by? The −fallacy button pulls up all 16 — pick the one that fits and it docks the point and tags the transcript right where it happened.
Or skip the button entirely — highlight the phrase yourself in the live subtitles and a "Fact-check this" shortcut pops up, running the exact same check on demand.
6 End of debate
When the clock runs out — or you end it yourself — everything lands on Results: final scores, the fact / wrong / fallacy tally for each side, a highlight reel of every moment that got flagged, and the full transcript, ready to export.
Turn on AI Judging and you also get an actual verdict — a points-based one that just does the math, or a substance-based one that reads the whole transcript and calls who made the stronger argument.
The six formats at a glance
| Format | Sides | Structure |
|---|---|---|
| Casual | Debater A vs B | Configurable rounds & timer. |
| Public Forum | Pro vs Con | Constructives · rebuttals · summaries · three crossfires · final focus. |
| Lincoln-Douglas | Affirmative vs Negative | 1-on-1 values debate; AC · NC with cross-ex · 1AR · NR · 2AR. |
| Policy (CX) | Affirmative vs Negative | Evidence-heavy; four constructives, four cross-exams, four rebuttals. |
| Parliamentary | Government vs Opposition | PMC · LOC · MG · MO · LOR · PMR. Points of information expected. |
| Mock Trial | Prosecution vs Defense | Openings · direct & cross exams per witness · closings · rebuttal. |
Privacy
Speech is turned into text first. By default your browser's own speech recognition does that — in Chrome and Edge it is run by Google or Microsoft, not by us. In High-accuracy mode, in Upload mode, and if your browser can't listen, short audio clips go through our server to Groq for transcription, and we don't keep them. The live checks then send only a few recent sentences and the topic to Groq — no audio and no names — and claims that look recent are looked up on Wikipedia from your browser. The AI judge and analysis do send the round's transcript and the names you entered. Nothing about a round is stored on our servers unless you sign in, enter a League code, or play an online room. If the AI service is busy, checks slow down or pause and the app keeps working in manual mode. The Privacy Policy has the full list.
When to overrule the AI
It will get things wrong sometimes — statistics, very recent events, and anything that hinges on exact wording are where it stumbles. It's tuned to stay quiet unless it's genuinely confident, but "genuinely confident" still isn't "certain." Watch the banner. If a call looks off, it probably is — undo is one click away in the tray, every time.