Chargement...
Chargement...
Under the hood
Seven steps, several specialized AIs, and checks at every stage. Here is what happens between a speech airing and the moment you can find it, cite it, and verify it.
The corpus, live
—
hours of video
—
videos
—
channels monitored
—
start of corpus
Processing pipeline
Every video follows exactly the same path, automated end to end, within a few hours of publication.
TV, radio, online media, parliamentary channels, official accounts: new videos are spotted every day.
✓ Sources hand-picked one by one
An AI reviews each spotted video and keeps only those where a political figure actually speaks.
✓ When in doubt, the video stays out
Full audio runs through our own speech recognition. Never platforms’ auto-captions. Every word is timecoded.
✓ Word for word, no summaries
The recording is split turn by turn: who speaks, from which second to which second.
✓ Each turn precisely timecoded
Each voice is matched to our personalities registry by cross-checking several clues: on-air introduction, context, program.
✓ Name shown only if confirmed
Each video gets one format among 8, each ~250-word segment one theme among 16.
✓ Same lists for everyone
An exact-words index, and a meaning index: each segment gets a digital fingerprint of what it says.
✓ Searchable by words and by idea
Ready to search,
cite and verify
Anatomy of a speech segment
A speech segment is not just text: it is a complete record, so you can find it from any angle and trace it back to the source.
Reliability
Attributing a sentence to the wrong person would be the worst error. Our rule: better an unidentified speaker than a wrong name.
Identity established by several consistent clues: the name is shown, with a link to their profile and speaking time.
Journalists, hosts, guests outside our registry: their role, and their name when given on air, without a personality profile.
If doubt remains, no name is invented. The sentence stays in the transcript, and identification can be completed later.
85 %
of political figures’ speaking time is attributed by name, with at least 80% confidence
Of which 78% are linked to a personality profile. The remaining 15% are left unnamed rather than risk an error, and the rate increases as our registry grows. Measured over about 7,000 hours of political speech, as of November 21, 2026.
In English
Everything French politicians say is transcribed in French. The English version of Dixipolis bridges the gap without touching the original.
1
Type a question or keywords in English: they are translated into French, then searched across the whole French corpus, by exact words or by meaning.
2
Excerpts, answers, fact-checks and articles are shown in English, translated as faithfully as possible, spoken register included.
3
Each quote keeps its French original (“Show original French”) and a link to the exact second of the video. A full transcript is translated in one click.
AI
Several AIs work in sequence, each on a specific task. None browse the Internet or rely on memory: everything comes from the transcripts.
1
The question is split into searches: which personalities, which period, which themes, which words, which idea.
2
Each search queries both indexes at once. Many candidate segments surface.
3
A second AI rereads each segment and discards those that do not really answer the question.
4
The answer is written only from the retained segments, each numbered and linked to its exact video second.
Fact-checking
Spot verifiable claims: figures, dated facts, comparisons.
Prioritize those that matter for the argument, of public interest.
Confront with official series (Insee — French statistics office, Eurostat) queried directly, then with a closed list of public sources and fact-checking newsrooms.
Judge taking into account the full sentence context and reasonable rounding.
Six verdicts, always sourced
A confidence level displayed
Below 70% confidence, no verdict is final: the line is marked “To verify”.
Everything is automated, so a transcription or attribution error is still possible. That’s why each segment links to the exact second in the video: the video prevails. Spotted an error? Report it, and we will fix it.
13,543
hours of video
62,571
videos
1,436
channels monitored
November 2010
start of corpus
Last 30 days: 1,455 hours and 2,826 videos added.