A Clip With Music Only: What Comes Back

The same 15 seconds, three times

One file, a C major chord held for 15 seconds, uploaded three times.

• First run: 33 characters, 6 words, 1.49 s
• Second run: 33 characters, 6 words, 1.41 s
• Third run: 33 characters, 6 words, 0.95 s

Identical text each time. So re-running will not save you here - the hallucination is as stable as a real transcript is. The wall clock moved by half a second across the three runs while the words stayed put.

Length is where it falls apart

We held the same chord for 5, 15, 30 and 60 seconds and sent all four.

• 5 s: 31 characters, 6 words, 0.83 s - "Thank you so much for watching."
• 15 s: 33 characters, 6 words, 1.49 s - "Thank you very much for watching."
• 30 s: 1,435 characters, 284 words, 11.99 s
• 60 s: 1,583 characters, 301 words, 41.00 s

The 30- and 60-second results are one sentence repeated until the file runs out - "This is the first time I've ever been to a place like this, where I've never seen a place like this..." over and over. Sixty seconds of music took 41 seconds to process and came back with 301 words that nobody said.

Not every kind of music behaves the same

Four kinds of sound, 15 seconds each, no speech in any of them.

• Three-note chord: 33 characters, 6 words, 1.49 s - "Thank you very much for watching."
• Single 440 Hz tone: 1,120 characters, 225 words, 9.61 s - "I'll see you in the next video, bye bye, bye bye..."
• Gated noise beat: 444 characters, 1 word, 7.94 s - one repeated "M-M-M-M-" token
• Bass with a vibrato melody: 34 characters, 7 words, 1.34 s - "Thank you very much for your time."

A steady tone was the worst of the four by a wide margin: 1,120 characters against 33. A steady chord and a melody both came back with a single polite sign-off.

Turning it down does not help

We changed the level on the same 15-second chord and re-sent it.

• -30 dB: 20 characters, 3 words, 1.23 s - "Thanks for watching!"
• -12 dB: 20 characters, 4 words, 1.00 s - "Thank you very much."
• +6 dB: 33 characters, 6 words, 1.06 s - "Thank you very much for watching."

Quieter music gets a shorter answer, not a truer one. There is no level at which the model admits it heard nothing.

One case where the music stays quiet

We put 13 seconds of the chord in front of 2 seconds of real Spanish speech - the shape of a clip with a music intro. The response was 13 characters and 3 words: "Hola a todos." The music added nothing at all. Real speech after music does not pick up a junk prefix, so an intro is not what breaks a clip. Music with no speech anywhere in it is.

How to tell it is fake

Two checks that worked today. First, characters per second: real speech ran 15 to 17 characters per second across a length ladder we ran today, while the 30-second music clip ran 47.8 - three times the words for the same length. Second, read it. Every music result today was one short sentence, or one short sentence repeated. If the whole transcript is a sign-off, there was nothing to transcribe.

Re-encoding will not shake it loose either. The same 15 seconds came back as the same 33 characters in MP3 at 64 kbps and in a 1.3 MB WAV. Dropping to 8 kHz did change it - 70 characters instead of 33 - because downsampling feeds the model different audio, not because it exposes the problem.