Five seconds of hiss, and two AIs recited me a sentence

In short

• Two transcribers (AIs that write down what they hear), on the sound alone, black picture.
• 6 videos out of 6: the two expected lines, word for word, from both sides.
• Cost: about $0.006 for 15 calls to Gemini, $0 for Whisper. If you check 100 videos: roughly $0.05 to $0.09.
• Rule: we only announce correct diction when both tools agree.

The hiss that talked

Hang on, listen to this. I made 5 seconds of background noise on a computer: a "shhhh" like a badly tuned radio, no voice, no breathing, nothing. I slip it into two transcribers. Both hand me back a sentence. A real one, with a subject and a verb.

• Gemini: "On part sur le deuxième plat. Ça s'appelle la fleur." ("We'll go with the second dish. It's called the flower.")
• Whisper: "Sous-titres réalisés par la communauté d'Amara.org" ("Subtitles made by the Amara.org community")

Right. A dish called the flower, I've eaten worse. But let's rewind.

Where it all started

In our "café" shot, two friends must say: "Tu viens avec moi demain ?" ("Are you coming with me tomorrow?") then "Oui, mais c'est toi qui conduis." ("Yes, but you're driving.") My mission: check that the video has them say it right, without listening to everything by hand.

The plan: a transcription AI writes down what it hears, I compare. Except our usual tool, Scribe, hosted by the fal platform (which runs AIs for pay), was blocked for lack of credit. I needed a reliable replacement.

Sound only, and two witnesses

First reflex: I cut the picture. Black screen, soundtrack kept. Some models burn subtitles into the picture, and a transcriber that sees them might copy them instead of listening. Real example: one wrote "Tu viens avec deamin.. ?" on screen while the voice did say "demain".

Second reflex: a single witness can lie, so I take two.

• Gemini, Google's AI (model gemini-3.8-flash), run twice per video with the instruction: word for word, no correcting or completing, "inaudible" if a word can't be heard, "no speech" if nobody speaks.
• Whisper, OpenAI's free software (the "small" version, one of the software's sizes), which runs at our place, offline.

Neither knows the expected text. Two witnesses who haven't talked to each other: if they say the same sentence, word for word, you can believe them. One alone can make things up.

Six videos, six times the same

I run everything on six videos of the same shot, from three families of video models: LTX 2.5 (in two versions of its text encoder, the part that reads the prompt), LTX 2.3, and MiniMax H3 (normal setting, then a fast mode that computes in 4 then 8 steps, i.e. fewer passes of computation). The numbers are versions.

For all six, both transcribers return exactly:

• "Tu viens avec moi demain ?"
• "Oui, mais c'est toi qui conduis."

Gemini gives the same text on both of its passes. Whisper adds timestamps: for LTX 2.5 (official encoder), the first line runs from 0.0 to 2.0 seconds, the second from 2.0 to 4.0.

Six out of six. Too good. I distrust myself: three checks.

Three traps to catch myself red-handed

1. Calibration. On 2 videos already transcribed on 26/10/2026 by Scribe, Gemini and Whisper find exactly its text.
2. Cut. I cut the sound at 2.2 seconds, right at the very start of the second line (Whisper dates it at 2.0 s on one of the two videos, LTX 2.3) (2 videos cut). Gemini only returns the first line, "Tu viens avec moi demain ?". It doesn't invent the continuation it "should" hear.
3. Pure noise. The hiss from the start, that is. It has nothing to do with the café videos, and that's the whole point: no voice in it. And here, both make things up.

So a single transcriber can "hear" words that don't exist. For Whisper, our hypothesis (not a measured fact): it probably saw that Amara sentence thousands of times at the end of subtitled videos during its training.

So why does the method still hold? On our videos, both tools wrote the exact words of the script, which neither knew. Inventing the same right sentence is very unlikely (our reasoning, not a measurement). And on noise, their inventions differ: the disagreement shows.

What it costs

Gemini. 15 calls: 12 transcriptions (6 videos, 2 passes each) and 3 checks (2 cuts, 1 noise). A "token" is a small piece of text or sound that the AI reads or writes: 7,191 read and 242 written in total. Google's displayed price: $0.75 per million tokens read and $3.75 per million tokens written until 31/12/2026 ($1.50 and $7.50 from 01/01/2027).

Calculation: about $0.005 for reading plus $0.001 for writing, i.e. about $0.006 for the 15 calls. Caution: the page gives a general rate; if Google bills sound differently, the total may change, but it would stay in the order of a few cents.

Whisper. $0: it runs on our server, already paid for.

If you check 100 videos: about $0.05 with one Gemini pass per video, about $0.09 with two passes (our calculation, at about 515 tokens read and 18 written per transcription), plus Whisper for $0. The price of the former tool, we don't have it: I'm not comparing.

What my two witnesses can't see

• Lip sync (does the mouth move at the right time?).
• Timbre, accent, emotion.
• Who says which line: the two friends could be swapped.

They only say: the words are right and understandable.

The rule, and the fifth glass of milk

A single transcriber can make things up. So we only announce correct diction when two independent tools agree, on the sound alone. That's worth 4 glasses of milk 🥛🥛🥛🥛. The fifth I'm saving for the day they stop hearing voices in the hiss.

Passages sources

Show the 8 sources

Réf. (S = source, R = in-house record)	Citation (langue originale)	URL
S1↑	Méthode : « Son seul : chaque v2-cafe.mp4 remonté sur une image NOIRE (ffmpeg) » ; Gemini gemini-3.8-flash, 2 passes par vidéo, consigne « mot pour mot, sans corriger, [inaudible] si besoin »	— (test maison)
S2↑1↑2↑3	Transcriptions Gemini des 6 vidéos (deux passes identiques), contrôles : coupe à 2,2 s (« Tu viens avec moi demain ? »), bruit pur (« On part sur le deuxième plat. Ça s'appelle la fleur. »)	— (test maison)
S3↑1↑2↑3↑4	Transcriptions Whisper avec horodatage ; bruit pur : « Sous-titres réalisés par la communauté d'Amara.org »	— (test maison)
S4↑	Whisper small, faster-whisper 1.2.1, local, processeur, langue fr, beam 5	— (test maison)
S5↑	Tarif de gemini-3.8-flash : 0,75 $ / 3,75 $ par million de jetons jusqu'au 31/12/2026, 1,50 $ / 7,50 $ ensuite ; page lue le 02/11/2026 vers 13:45, heure universelle	https://ai.google.dev/gemini-api/docs/pricing
R1↑1↑2↑3	Résultats : 6 vidéos, concordance exacte des deux transcripteurs ; étalonnage sur 2 vidéos transcrites par Scribe le 26/10	— (test maison)
R2	Tableau des transcriptions des 6 vidéos	— (test maison)
R3↑1↑2↑3↑4	15 appels Gemini, 7 191 jetons lus, 242 écrits ; Scribe bloqué (solde fal à 0 $) ; exemple « deamin »	— (test maison)