▶TranscriptReel

Voice Isolation: What Four Cleanup Filters Do to the Words

2026-11-11
Short answerThe noisy file did not need cleaning. A 45-second clip mixed with pink noise at the level of the voice returned all 150 words on the first upload, plus two invented ones. Spectral denoise came back with 148 words and "four" written as "4". Dialogue enhancement dropped the same three-word phrase. A band-pass filter printed "fix that" twice.

What we actually fed it

Forty-five seconds of one voice reading about a microphone and an air conditioner, written and synthesized for this test so the correct text is known before anything is uploaded. The clean version is 3,969,078 bytes as 44.1 kHz mono WAV.

Then we added pink noise at amplitude 0.11 across the whole 46 seconds and summed the two tracks without normalizing. The result is 3,969,078 bytes of WAV that sounds, on headphones, like someone talking in a room with a fan running. The noise is not subtle. It sits at roughly the level of the speech, which is louder than most real recordings manage.

A laptop and a small microphone on a wooden desk beside a window with net curtains and daylight
One voice, one fan, five ways out.

The unfiltered mix won

Sent with no processing at all, the noisy file returned 800 characters and 152 words. Against the clean file's 787 characters and 150 words, that is every real word plus two that were not there. The alignment shows a single insertion at the end: "didn't work".

Read that again. The last sentence of the clean recording ends at "closer to fix that, and it". The clean upload was cut off there. The noisy upload finished the sentence and added two words that are not in the script. The transcript is more complete than the clean one and less accurate at the same time.

So the first result of this test is that noise at speech level did not break anything. The recognizer has been trained on exactly this, and a fan is not what stops it.

Four filters, four different failures

We then cleaned the same mix four ways, each one a filter you would find in a normal editor or reach for after a search for voice isolation.

Every filtered version is worse than the unfiltered one on word count. Three of the four lost the same two words. That tells you where they sit: at the end of the recording, in the quietest part, right after the sentence's main clause.

The band-pass result is the interesting one, because it did not lose the words. It duplicated the tail of the sentence. Stripping everything below 300 Hz and above 3.4 kHz removed the noise. It also removed the cues the model uses to decide a sentence has ended, so it kept going and re-emitted what it had already said.

Overlapping word counts for five versions of the same clip, with the noisiest version on top
152, 148, 148, 150, 148. The unfiltered mix holds the most words.

Why guitar-pedal thinking fails here

A denoiser is built to make audio pleasant. It removes the fan, the hum and the room, and it measures success by whether a person prefers the result. A recognizer is not a person and does not care about pleasant.

The model reads a spectrogram, and what it wants from that spectrogram is contrast between the speech and everything else. A denoiser that removes the noise also removes the low-level detail near the speech, because it cannot tell the difference between a quiet consonant and a quiet anything-else. The gate does it even more bluntly: anything under the threshold becomes digital silence, and consonants live under thresholds.

A gate at 0.03 might look gentle on a meter. On this file it turned the tail of a sentence into nothing, and the model, faced with nothing, either stopped or invented.

The one number that explains the pattern

Count how much each version actually had to work with. The clean file is 3,969,078 bytes, the mix is the same size, and all four filtered versions were written back at 192 kbps, 1,081,512 bytes each. Every filtered version has the same amount of audio and the same encoding.

So the difference is not bytes and not duration. It is entirely what the filter did to the waveform in between. That is the cleanest way to see it: four files, one size, four different transcripts, and the only variable is the filter chain.

What did move, and what would have moved anyway

Two of the five versions wrote "four" as "4". The clean file and the raw mix both wrote it out. That is not a word error, it is a formatting decision, and it appeared only after the audio had been through a filter. Our earlier tests on numbers show the same thing happening without any filtering, so I would not blame the denoiser for it.

The dialogue enhancement added a comma after "there". Punctuation on this endpoint is already loose, and a comma appearing or disappearing between two uploads of the same audio is normal. We have seen the character count move by five across five re-encodes of an identical bitstream.

What we would tell someone with a bad recording

Send it first. Before you clean anything, upload the file you have and read the result. If the words are all there, the cleaning is a downgrade, and this test shows it costing you three words out of 150.

If the words are not all there, then the noise really is in the way, and the fix is not a filter. It is a different take, a closer microphone, or a quieter room. None of those are available after the fact, which is why the honest answer is that a bad recording stays bad.

The narrow case where a filter helps is when the noise is a single steady tone, like a mains hum or a fridge, and the speech is well above it. A notch or a gentle high-pass there removes a specific frequency without touching the band speech lives in. Broadband cleaning is the thing to avoid.

A screenshot-style comparison of five transcript excerpts, one per filter chain
Same 45 seconds, five cleanups, five different endings to one sentence.

Where your file goes

All five uploads went to /api/transcribe over TLS, were processed once in memory, and were discarded when the response came back. Nothing is stored between uploads. That is why sending the same file twice is a valid experiment and not a cached answer: the second request had nothing to read but the bytes.

Next step

Got a Reel you want as text? Download the Instagram transcript as TXT, SRT, or VTT →

Affiliate link

If the noise really is a single steady tone, take it out by frequency rather than by cleaning everything. Wondershare UniConverter runs on your own machine and lets you cut a band without resampling the rest of the recording, which is the one cleanup that did not cost words here. It does not transcribe - the text comes back from this page.

FAQ

Should I remove background noise before transcribing?

Try it without first. Our 45-second clip with pink noise at the level of the voice returned all 150 words plus two invented ones. Every one of four cleanup filters returned fewer words, and three deleted the same phrase at the end of a sentence.

Why did a noise gate make the transcript worse?

A gate turns everything under its threshold into silence, and quiet consonants sit under most thresholds. At 0.03 with a 6:1 ratio, our file lost the words "and it" at the end of a sentence, dropping from 152 words to 148.

Does a band-pass filter help a noisy recording?

It removed the noise and duplicated a sentence. High-pass at 300 Hz plus low-pass at 3.4 kHz returned 790 characters and 150 words. The sentence ending, though, changed from "to fix that, and it" to "to fix that. fix that."

Which cleanup filter changed the words least?

The raw mix, with no filter at all. It returned 152 words against 148 for three of the four filters and 150 for the band-pass. The unfiltered upload was also the only version that finished the last sentence.