An English Speech Model Invented Japanese Dialogue, and Its Confidence Gate Caught It

The English speech model marked 23 of 27 Japanese lines as doubtful, and it was right.

That result came from a transcription attempt I tested while measuring the finished film. The transcription itself was not usable. The confidence gate was.

That distinction matters. A failed output does not mean every part of the process failed. In this case, the model could not give me dialogue I could trust, but it could tell me that I should not trust what it had produced.

I kept the failed attempt for one reason: I did not want to repeat it later and mistake a formatted transcript for evidence.

The transcription was outside the working measurement chain

My finished measurement process uses four programs in a fixed order.

The first measures duration, cuts, and sound level.

The second builds numbered contact sheets with one frame from every shot and the minute attached.

The third extracts the subtitles written on screen, crops them, and assembles them into readable sheets.

The fourth checks the shot count against twenty cuts, showing the frame before and the frame after each one.

All four run on their own. None asks for input.

The transcription attempt was different. It listened to the audio and tried to turn the spoken voices into text. That sounds like another measurement step, but the result did not belong beside the other four. The voices were Japanese. The speech model was English. It invented dialogue.

I wrote:

las voces están en japonés y
el modelo en inglés alucina. Su propia puerta de confianza marcó 23 de 27 líneas como
dudosas y acertó.

The useful English reading is simple: the voices are Japanese, the model is English, and it hallucinates. Its own confidence gate marked 23 of 27 lines as doubtful, correctly.

That is the whole diagnosis supported by the test. I do not have a basis here for a wider technical explanation of why the model behaved that way. I only have the language mismatch, the invented transcription, and the confidence result.

Those three facts were enough to reject the output.

The confidence gate did the job the transcript could not

I wanted the transcript to tell me what was said. It failed at that job.

The confidence gate answered a different question: which lines should be treated as doubtful?

It marked 23 of the 27 lines. The judgment matched what I heard. That made the gate the useful part of an otherwise unusable pass.

I would not promote those flagged lines into subtitles, dialogue records, or evidence about the finished film. A line that looks grammatically complete is still not trustworthy when the system producing it says its own confidence is low and the listening result is wrong.

The temptation is to rescue the output by correcting a few obvious errors. I was not dealing with a missed word here or a confused name there. I was dealing with an English model hallucinating on Japanese voices.

That changes the decision. This was not a clean transcript with minor damage. It was a result to reject.

On-screen text and spoken dialogue were separate jobs

The working measurement chain already had a way to capture the subtitles written on screen.

That program did not listen to the voices. It took the written subtitles from the picture, cropped them, and placed them onto readable sheets.

The transcription attempt listened to audio.

Those are not interchangeable inputs. One reads text already present in the finished frames. The other tries to infer words from speech. In this film, the first process remained part of the final measurement sequence. The second was a test that does not work.

That separation kept the bad transcript from contaminating the useful subtitle record.

It also gave me a cleaner standard for acceptance. The subtitle sheets came from text visible on screen. The transcript came from Japanese voices interpreted by an English model. Once the confidence gate marked 23 of 27 lines as doubtful, there was no reason to treat those two outputs as equally dependable.

Confidence is a gate, not a repair

The gate did not turn the invented dialogue into correct dialogue.

It did not replace the 23 doubtful lines. It did not verify the remaining lines. I only know that its doubtful judgment was correct.

That is narrower than saying the model corrected itself. It did not.

What worked was the stop condition. The system produced an output, measured its confidence, and gave me enough warning to keep that output out of the final measurement chain.

That is valuable because autonomous programs do not ask me to inspect every intermediate decision. The four working measurement programs run without requesting anything. If I add another autonomous process, I need a clear way to decide whether its result belongs in my finished measurements.

For this transcription attempt, confidence supplied that decision. Twenty-three doubtful lines out of twenty-seven was not a result to polish. It was a result to stop using.

I kept the failure documented

I wrote that the transcription attempt does not work so nobody repeats it.

Keeping that warning is part of the result.

Without it, I could return to the same idea later, see a speech model produce a full set of lines, and waste time testing the same route again. The failed transcript might even look usable before comparison because it arrives as text. The confidence gate is what exposed the problem inside the process itself.

I preserved both halves.

The English model invented dialogue from the Japanese voices.

The model’s own confidence check caught the failure by marking 23 of 27 lines as doubtful.

I trust the second result. I reject the first.

That is not a contradiction. They are different outputs answering different questions.

The acceptance rule

For this film, an audio transcript from that English model is not part of the finished measurement process.

The accepted sequence remains the four autonomous programs: measure duration, cuts, and sound level; build one-frame-per-shot contact sheets; extract the written subtitles from the screen; and calibrate the shot count against twenty before-and-after cut pairs.

The rejected path remains documented beside them.

That is what “caught it” means here. The confidence gate did not solve Japanese transcription. It prevented invented lines from being accepted as a measurement of the film.

Today, take one non-English clip you have already verified, run it through your speech model, and compare every low-confidence line against the known material before you use the transcript in your own work.

---

You just watched one shot come apart, decision by decision. None of them landed there by luck.