Home / Music Stress Test
Can Background Music Change an Instagram Transcript?
Yes, but not in every mix: in our controlled 15-word sample, the same Whisper tiny English model made 0 word errors on the clean voice and at -30 dB speech-to-music ratio, 1 error at -35 dB, and 16 word-level edits at -40 dB. The clean voice is an original stem retained before mixing, not the output of a source-separation product.
Hear the test inputs
We generated one English sentence with a local text-to-speech voice, then overlaid a simple, lyric-free three-tone pattern. The sentence, not a creator's Reel or caption, is the reference:
The yellow notebook is beside the blue lamp. Please send the revised draft before Friday.
What the recognizer returned
| Input | Word-level edits vs. reference | Observed output |
|---|---|---|
| Original voice stem | 0 | All 15 reference words matched. |
| Music at -30 dB | 0 | All 15 reference words matched. |
| Music at -35 dB | 1 | The final word changed from “Friday” to “writing.” |
| Music at -40 dB | 16 | Multiple substitutions and insertions; the output no longer preserved the sentence reliably. |
The edit count is the minimum number of word insertions, deletions, and substitutions needed to turn the recognized words into the reference after case and punctuation are ignored. It is not a percentage accuracy claim. At -40 dB, insertions make the edit count larger than the 15-word reference.
The -35 dB output was: “The yellow notebook is beside the blue lamp. Please send the revised draft before writing.” The -40 dB output was: “the yellow-nubel of his beside him, and the lamp, and the basement, and the revised draft of the morning right at me.” These are raw model words with only the leading space removed. The original voice and -30 dB mix both returned the reference sentence, aside from capitalization.
How we made and checked it
On October 14, 2026 (Los Angeles time), we used the local Microsoft Zira Desktop voice to render the reference as 16 kHz, mono, 16-bit PCM. We generated music from 220, 277.18, and 329.63 Hz tones with a 2 Hz pulse, then mixed it with the retained speech stem at measured speech-to-music RMS ratios of -30, -35, and -40 dB. We normalized each mix to avoid digital clipping. Each input went through Xenova/whisper-tiny.en with the same transcription call. The four WAV files above let you listen to every tested input.
The -30 dB result matters as much as the failures: this particular model recovered all 15 words despite a strong tonal bed. The -35 dB mix changed one word, and -40 dB changed many. That is a boundary observed for one short synthetic voice and one artificial instrumental pattern, not a threshold to apply to every Reel. Sung vocals, percussion, another speaker, different languages, accents, codecs, or a real platform download may behave differently.
No source-separation model was run. Because we created the mix, we already had the clean speech stem; using that stem is an ideal reference, not evidence that a tool can recover it from an unknown Reel. We also did not submit these files to GetTheScript or any competing service.
What this adds to a tool claim
When we opened GetTheScript's Instagram transcript page, it offered a paste-link workflow and said its audio-first engine transcribes speech even under music and effects. That page did not provide a matched, downloadable clean-versus-mixed audio sample or a word-error table. Our table supplies those for this model only; it does not rank the two services or validate the competitor's result.
If you need text from a music-heavy public Reel, run the link, then listen back while checking names, numbers, and quotations. Compare the output against the actual speech before publishing it. The earlier music-path check explains why XAPN does not first extract a clean vocal stem. A finished transcript and a trustworthy quotation are different things.
The controlled samples explain a boundary, not every Reel. Use the main workflow to turn an Instagram Reel into text when you have a public link to process. Judge that draft against its own source audio.