How to settle a word you cannot make out in a noisy recording

Last updated 2026-09-15 · 8 min read

Tell Vizard Agent which two words it might be. It cleans the track gently — a heavy denoise removes the consonants that distinguish them — re-transcribes a longer stretch so the surrounding sentence constrains the answer, and then compares the two candidates directly rather than transcribing the phrase again from scratch.

What is the short version?

A transcript that is ninety-nine per cent right can still contain one word that changes the meaning of the sentence it sits in. Playing the passage again does not help much, because by now you have already heard it the way you expect to hear it, and each repetition makes that reading firmer.

Tell Vizard Agent which word and which two it might be.
Ask for a gentle clean rather than a heavy one.
Ask it to transcribe a longer stretch around the word.

What do you need before you start?

The recording and your best guesses. Vizard Agent works far better on "is it this or that" than on "what is this", because two candidates turn an open question into a comparison — and a comparison can be settled.

The recording. The original, not a re-encode.
Where the word is. Roughly is fine.
The two candidates. Both spellings.
The sentence around it. As much as you have.
The language. Especially if more than one is spoken.

What do you type into Vizard Agent?

Give Vizard Agent both words. The instinct is to ask what the word is, but that is the question the first transcript already got wrong, and asking it again in the same open form tends to produce the same wrong answer with the same confidence.

Prompt

Around the twenty-second mark the phrase starts "on some ..." and I cannot tell whether the word is "hollows" or "hillocks". Clean the noise off the track gently — not so hard that it damages the speech — then transcribe a longer stretch around it and tell me which of the two it is.

Variants worth knowing:

"Whether it is X or Y." Turns it into a decision.
"Gently, not so hard it damages the speech." The important caveat.
"A longer stretch around it." Context does the work.

What does Vizard Agent actually do?

Here is the sequence on a real recording where a single word was in dispute. Notice that the first clean was too aggressive and had to be redone more gently — which is the lesson of the whole job.

Transcribes the speech with timings.
Re-listens to the short phrase and confirms the language.
Cleans the background noise off the track.
Re-transcribes the cleaned audio.
Listens to the cleaned track and narrows the disputed word.
Redoes the clean gently, with the voice boosted instead.
Re-transcribes after the gentler processing.
Takes a longer stretch of the original and transcribes that.
Cleans the longer stretch gently and transcribes it again.
Compares the two candidate words directly on the cleaned audio.

Step six is the finding worth keeping. Noise reduction works by removing what does not look like speech, and the quiet consonants at the ends of words are exactly what it mistakes for noise — so the aggressive pass makes the recording cleaner and the word less identifiable.

Step eight is the other half. A word heard alone is a coin toss; a word heard inside a sentence is constrained by grammar and sense, so transcribing thirty seconds settles what transcribing two seconds cannot.

Step ten is why the candidates matter. Vizard Agent is not being asked to produce the word from nothing — it is being asked which of two fits the audio better, and that is a much easier question to answer honestly.

What does the result look like?

An answer with a reason, or an honest failure. Vizard Agent tells you which candidate the cleaned audio supports and gives you the corrected transcript as a file — and if the recording genuinely cannot decide, it says that rather than picking one.

When does this not work well?

Some words are simply not recoverable. If the noise sits at the same frequencies as the speech, or the speaker mumbled straight through the word, then no amount of processing will put back information the microphone never captured in the first place, and Vizard Agent will tell you so.

Noise over the speech. Same frequencies, no separation.
Clipped or distorted audio. The detail is gone, not hidden.
A mumbled word. Even a person in the room might not know.
Heavily compressed uploads. The re-encode removed the fine detail.
Two speakers at once. Separation first, then transcription.

How do you fix a result that came back wrong?

Add context rather than repeating the question. If Vizard Agent picked the word you do not believe, telling it what the conversation is about, or offering a third candidate spelling, gives it something genuinely new to work with — whereas asking again gets you the same answer.

"Neither of those fits." A third candidate, or a wider stretch.
"The clean sounds worse." A gentler pass, or none at all.
"It is a proper noun." Say so; it changes what is plausible.
"They are talking about X." Subject matter narrows the options.

How does Vizard Agent compare to doing it yourself?

By hand you loop the two seconds twenty times, and by the tenth you can hear whichever word you last thought of. Then you run a denoise, it sounds dramatically cleaner, and the word is now less recoverable than before.

	By hand	Vizard Agent
The question	What is this word?	Which of these two is it?
Denoise	Set for how clean it sounds	Set for how well it transcribes
Context	The phrase on loop	A longer stretch transcribed
Your expectation	Decides what you hear	Not in the loop
The outcome	A guess you will re-check	An answer or an honest no

Common questions

Why does denoise make it harder? Because quiet consonants look like noise. Vizard Agent cleans gently for this reason.

Do I have to give two options? No, but it helps a great deal. Vizard Agent can compare candidates directly.

How long a stretch should it transcribe? Enough to give the sentence. Vizard Agent widens it when the short version fails.

Can it tell me how confident it is? Yes. Vizard Agent says when the audio does not support either candidate.

Will it fix the whole transcript? Yes, if you want. Vizard Agent corrects the word and gives you the file.

What if two languages are mixed? Say so. Vizard Agent confirms the language of the phrase before transcribing it.

Can it clean the audio for delivery too? Yes, but that is a different setting. Cleaning for transcription is gentler.

What about background music? Vizard Agent can separate that first, which often helps more than denoising.

Does the original file get changed? No. Vizard Agent works on copies and gives you the transcript.

What if the speaker mumbled? Then it may be unrecoverable, and Vizard Agent will say so.

Can I send just the clip with the word in it? You can, but send more. Vizard Agent uses the surrounding speech to decide.

Will it guess if it cannot tell? No. An honest "I cannot tell" is more useful than a confident wrong word.

Can I get the timings too? Yes. Vizard Agent transcribes with timings by default.

What if it is a name? Tell Vizard Agent it is a name — that changes which spellings are plausible.