How do I get cleaner audio before sending it to be transcribed?

This page may contain affiliate links from which Blur Studio earns commissions.

This is a cheap, genuinely obvious step that a surprising number of people skip entirely: cleaning the audio before transcription instead of after the fact. Transcription quality, whether it's a human service or an automated one, depends heavily on how clearly the words are actually captured in the recording. Noisy input produces more guessed words, more misheard names, and more manual correction on the back end, no matter how good the transcription tool is. For phone recordings in particular, see making hard-to-hear recorded customer calls clearer.

Why clean input matters more than people expect

A transcription model or a human transcriber is working from exactly what's in the recording. If background noise is competing with the speech, both have to guess at words that are partially obscured, and both will guess wrong more often as the noise gets louder or more inconsistent. Fixing this before transcription, rather than fixing the transcript afterward, is almost always less work overall, since you're correcting one audio file once instead of proofreading an entire transcript line by line.

How to clean the audio first

1. Test on one short clip before committing to a whole recording. Run a representative minute or two through ElevenLabs' Voice Isolator and transcribe both the original and the cleaned version of just that clip, then compare the two transcripts directly for accuracy.
2. If the cleaned version transcribes noticeably better, run the full recording through the isolator before sending it anywhere for transcription. It accepts most common audio and video formats, up to 500MB and 1 hour per file.
3. Send the cleaned file to your transcription service, not the original. This sounds obvious but is the step people actually forget, especially if cleanup and transcription happen through two different tools or two different people.
4. Budget the cost in. Voice Isolator runs on credits, roughly 1,000 credits per minute processed, which is a small cost relative to the time saved correcting a bad transcript afterward.

A worked example: a multi-speaker panel recording

Say you recorded a panel discussion, three or four people, a room mic rather than individual lapel mics, with some audience shuffling and a slightly echoey room. This is a harder case than a clean one-on-one call, since the noise isn't just background hum, it's room acoustics plus multiple voices at varying distances from the mic. Here's the sequence that actually works:

1. Pull a 2 to 3 minute test clip that includes at least two different speakers, not just the loudest one, since that's where transcription accuracy differs most between a clean and a noisy version.
2. Transcribe the raw clip first and note where it struggles, specifically: does it misattribute who's speaking, drop words during overlapping speech, or just mishear individual words? These are different problems with different fixes.
3. Run the same clip through Voice Isolator and transcribe the cleaned version. If speaker misattribution improved along with word accuracy, cleanup is solving a real problem here and it's worth processing the full recording.
4. If speaker misattribution is still a problem after cleanup, that's a different issue, probably better addressed by your transcription tool's own speaker-diarization settings than by audio cleanup, since isolation cleans the audio, it doesn't tell a transcription tool who's talking.

Worth noting separately for a panel or any multi-speaker setup: individual lapel or clip mics for each speaker, if that's an option on future recordings, solve a large part of this problem before you ever get to a cleanup or transcription step, since each voice is captured cleanly and separately rather than all blended together on one room mic. It's not always practical, a live panel in front of an audience is a different setup problem than a scheduled interview, but it's worth considering for anything you record regularly with the same small group of speakers, since the upfront equipment cost is usually small compared to the ongoing cleanup and correction time a noisy room-mic recording creates every single time.

This test-first approach matters more for multi-speaker recordings than simple ones, since the value cleanup adds genuinely varies by what's actually wrong with the recording, and a 2-minute test tells you that before you commit to processing (and paying for) the whole file.

Does the transcription tool itself already handle noise well?

Modern speech-to-text models have gotten genuinely better at handling imperfect audio on their own. ElevenLabs' own Scribe v2, for instance, publishes word-error rates under 5% across dozens of languages, with its best-case accuracy down near 2%, even without a separate cleanup pass. That doesn't mean cleanup is pointless, it means the gap it closes is narrower than it used to be for moderately noisy audio, and widest for genuinely hard cases: heavy, inconsistent background noise, multiple overlapping speakers, or a recording environment with real acoustic problems like echo. The practical takeaway is the same as the test-one-clip advice above: don't assume you need the cleanup step, and don't assume you don't. Test it against your actual recording and let the comparison tell you, rather than treating cleanup as either always necessary or always skippable.

When this step isn't worth it

If your recording is already reasonably clean, a quiet room, a decent microphone, one speaker at a time, this extra step adds cost and time for little real benefit. The test-one-clip approach above is exactly how to tell quickly: if the original clip already transcribes accurately, skip the cleanup step for the rest of the file and save the credits. This isn't a step to apply by default to everything; it's specifically for recordings you already suspect are noisy enough to cause real transcription errors.

Common mistakes worth avoiding

• Cleaning up audio after transcription instead of before. Running isolation on audio you've already transcribed doesn't help the transcript you already have; it only helps if you re-transcribe the cleaned version. Always clean, then transcribe, in that order.
• Processing the whole recording before testing a short clip. This is the single easiest way to spend credits you didn't need to. A 2-minute test tells you whether cleanup helps at all before you commit to the full file.
• Assuming cleanup fixes name and terminology errors. If your transcript consistently misspells a product name, a client's name, or technical jargon, that's usually a vocabulary problem, not a noise problem; check whether your transcription service supports a custom glossary or vocabulary list, which solves this more directly than audio cleanup ever will.
• Treating cleanup as a blanket rule rather than a per-recording decision. Audio quality varies by who recorded it, where, and on what device. A rule like "always clean up interview recordings" is less useful than actually checking each batch, since some will need it and some genuinely won't.

Roughly what this costs at volume

If you're transcribing regularly, say 20 hours of recordings a month, and roughly half of them are noisy enough to be worth cleaning first, that's about 10 hours of processing at 1,000 credits a minute, around 600,000 credits a month. That sounds like a lot until you weigh it against the alternative: an hour of recording with a rough transcript might need 15 to 30 minutes of manual correction, so 10 hours of noisy recordings could mean 2.5 to 5 hours of someone's time spent fixing transcripts by hand every month. Measured against that, the processing cost is a small, predictable line item against a much larger and less predictable cost in staff time. The math gets even more favorable the more recordings run through the same pipeline, since the correction time saved scales with volume while the per-minute processing cost stays flat and predictable.

How I'd approach this

For an occasional recording, testing one clip first and deciding from there takes a few minutes and is something anyone can do themselves. If your business regularly sends recordings for transcription, interviews, meetings, customer calls, and quality is inconsistent depending on who recorded what and where, I'd build a step into that pipeline that automatically checks and cleans audio before it ever reaches the transcription service, so the decision isn't made manually, clip by clip, every time. A well-built version of this also keeps a simple record of which recordings got the cleanup pass and which didn't, so if a transcript ever looks suspicious later, you know immediately whether it came from cleaned or raw audio rather than having to guess, instead of relying on someone's memory of what they ran it through weeks earlier. For calls specifically, transcribing every call automatically removes the per-recording decision.