How to Transcribe Audio With Strong Accents or Background Noise

You ran a recording through a transcription tool and got back something close to nonsense: names mangled, half the sentences marked as inaudible, two speakers merged into one run-on paragraph. The problem is almost never that the speaker's accent is "too strong." It is that the audio reaching the model is competing with a café, an air conditioner, a phone line, or a second person talking over the top.

Learning how to transcribe audio with strong accents or background noise is mostly a preparation problem, not a software problem. Fix the signal first, give the model the vocabulary it cannot guess, and pick the right language variant. The steps below start with free manual methods so they work regardless of which tool you finish in.

Step 1: Diagnose what is actually wrong with the recording

Before touching a setting, listen to sixty seconds with headphones and write down which of these you hear.

Constant broadband noise. Air conditioning, fans, road hum, hiss from a cheap preamp. This is the easiest to fix and the least damaging to accuracy once removed.

Intermittent noise. Door slams, cutlery, keyboard clicks, a dog. Harder to remove cleanly, but it only damages the words it overlaps.

Room reverb. The speaker sounds far away, as if in an empty hall. This is the single hardest condition to recover from, because the reflections smear the consonants that carry meaning.

Overlapping speech. Two or more people talking at once. No model handles this well; it needs to be split, not cleaned.

Low or wildly uneven levels. One speaker near the mic, another across the table.

Accent is a separate axis. A strong regional or second-language accent recorded cleanly usually transcribes well. The same accent recorded in a noisy room falls apart, because the model has fewer acoustic cues to fall back on. If you have both, treat the noise first and you will often find the accent problem largely disappears.

Step 2: Normalise levels before you touch anything else

Quiet audio forces every later processing step to amplify noise along with speech. Fix the level first.

In Audacity, which is free on Windows, Mac and Linux:

1. Open your file with File > Open.
2. Select all with Ctrl+A (Cmd+A on Mac).
3. Go to Effect > Volume and Compression > Normalize.
4. Tick Normalize peak amplitude to and set it to -1.0 dB. Leave Remove DC offset ticked.
5. Click Apply.

If one speaker is far louder than another, add gentle compression afterwards: Effect > Volume and Compression > Compressor, threshold around -18 dB, ratio around 3:1. Do not push the ratio higher; heavy compression pulls up room noise between words and makes things worse.

Step 3: Reduce background noise without destroying the voice

Audacity's noise reduction works from a sample of the noise alone, so you need a stretch with no speech.

1. Find one to two seconds where nobody is talking but the noise is present. The gap before someone starts speaking is usually best.
2. Select just that stretch.
3. Go to Effect > Noise Removal and Repair > Noise Reduction, then click Get Noise Profile. The window closes.
4. Select the whole track with Ctrl+A.
5. Reopen Effect > Noise Removal and Repair > Noise Reduction.
6. Set Noise reduction to about 9 dB, Sensitivity to 6.00, Frequency smoothing to 3.
7. Click Preview, listen, then Apply.

The mistake almost everyone makes is setting noise reduction to 30 dB or more. That produces silence between words and a watery, robotic voice during them, and transcription accuracy drops sharply because the artefacts destroy exactly the high-frequency detail that distinguishes similar consonants. Start low. If you can still hear noise but the voice sounds natural, you are in the right place.

For hum specifically, try Effect > EQ and Filters > Filter Curve EQ and roll off everything below about 80 Hz. Human speech has almost nothing useful down there, and removing it cleans up rumble at no cost.

Step 4: Split overlapping speakers into separate passes

If two people talk over each other, no amount of cleanup helps. You have two options.

If you recorded on separate tracks (two lapel mics, a multitrack recorder, a remote interview platform that saves per-participant files), export each speaker as their own file and transcribe them separately. Merge the two transcripts afterwards using the timestamps. This is by far the most accurate route, and it is worth changing how you record just to get it.

If you only have one mixed track, isolate the worst crosstalk sections manually. Cut the recording at the points where the overlap starts and ends, transcribe the clean sections normally, and hand-type the overlapping few seconds yourself. Overlaps are usually shorter than they feel.

For interviews where you know the audio is difficult, recording each participant separately from the start is the highest-leverage change you can make. Our guide to transcribing webinar recordings covers the same principle for multi-speaker sessions.

Step 5: Choose the correct language variant, not just the language

This is the step most people skip, and for accented speech it matters more than any audio processing.

Speech models are trained on regional variants, and selecting the right one changes which pronunciations the model expects. If your speaker has a strong Indian, Scottish, Nigerian, Singaporean or Irish accent, selecting the matching English variant rather than generic English or US English will noticeably improve results. The same applies across languages: Latin American Spanish versus European Spanish, Brazilian versus European Portuguese, Cantonese versus Mandarin.

Two related cases to watch:

Code-switching. If a speaker moves between two languages mid-sentence, set the primary language to whichever dominates and expect the other to come through phonetically. Some tools handle mixed-language audio better than others; test with a short clip before committing a long file.

Non-native speech. For a French speaker talking English, set the language to English, not French. Accent variant selection helps, but the language must match the words being spoken.

When you upload to Tapescribe you pick the language and variant before processing, so it is worth spending ten seconds getting this right rather than accepting the default.

Step 6: Add custom vocabulary and name hints

Every transcription model fails on the same category of word: proper nouns, product names, industry jargon, acronyms, and anything it has not seen in training. These are not accent failures; the model simply has no candidate that fits the sound.

Before you transcribe, write a list of:

• Every speaker's full name, spelled as you want it to appear
• Company, product and brand names
• Technical terms and acronyms specific to the topic
• Place names, especially local ones
• Any recurring phrase unique to this recording

Feed that list into your tool's custom vocabulary or keyword hints field. This single step often fixes more errors than all the audio processing combined, particularly for accented speakers whose pronunciation of a name is further from the model's expectation.

If your tool has no vocabulary field, do it afterwards with find and replace. Transcribe once, scan for the consistent misspelling the model chose, then replace all instances at once. Models are consistent in their errors, which makes this fast.

Step 7: Run the transcription and review the low-confidence sections

Upload the cleaned file and transcribe. When it comes back, do not read it start to finish; go straight to the places errors cluster.

Errors concentrate around the first few seconds of each speaker turn, around the noise events you noted in Step 1, at the ends of sentences where speakers trail off, and on any word from your vocabulary list. Play back with the transcript synced to audio so you can jump to a timestamp rather than scrubbing.

Tools that align the transcript to playback make this dramatically faster. Tapescribe returns timestamped output alongside subtitle formats, so you can click a line and hear exactly that moment. If you are working from a video source, the same timestamp approach applies; see our walkthrough on transcribing a YouTube video with timestamps. You can try it on a difficult recording to see how much your cleanup actually bought you.

Realistic accuracy expectations by condition

Setting the right expectation saves you from chasing an impossible result.

Clean audio, close mic, one speaker, any accent. Expect very high accuracy with only proper nouns needing correction. Custom vocabulary handles most of it.

Clean audio, strong accent, technical vocabulary. Good accuracy on ordinary words, poor on jargon until you supply a vocabulary list. This is the case where hints matter most.

Moderate steady background noise, close mic. After the cleanup in Steps 2 and 3, expect results close to clean audio. Steady noise is the friendliest kind.

Heavy reverb, distant mic. Expect meaningful degradation that processing cannot fully repair. Budget review time, and consider re-recording if the content is important.

Overlapping speakers on one mixed track. Expect the overlap sections to be unusable regardless of tool. Split or hand-type them.

Phone or compressed VoIP audio. Narrow bandwidth removes the high frequencies that distinguish consonants, so accented speech suffers more here than anywhere else. Vocabulary hints help disproportionately.

Common problems and how to fix them

Everything comes back as one wall of text with no speaker separation. Turn on speaker diarisation before transcribing rather than after. If speakers still merge, they are probably at similar levels through the same mic; separate recordings are the real fix.

The transcript is riddled with plausible but wrong words. This is a vocabulary problem, not a noise problem. Build the hint list from Step 6 and re-run.

Long stretches marked inaudible. Check your noise reduction setting. Over-aggressive reduction removes quiet speech entirely. Go back to the original file and redo Step 3 with a lower dB value.

Accuracy dropped after cleanup. You processed too hard. Transcribe the raw file as a control and compare. If raw wins, dial every effect back by half.

The first words of each sentence are missing. Some recorders use noise gates that clip speech onsets. Disable the gate at the recording stage; it cannot be recovered afterwards.

Voice notes from messaging apps transcribe badly. These are heavily compressed by design. Our guide to transcribing WhatsApp voice notes covers the format-specific handling.

Frequently Asked Questions

Does a strong accent really reduce transcription accuracy?

Less than most people assume, provided the recording is clean. Modern speech models are trained across a wide range of accents, so clearly recorded accented speech usually transcribes well. What actually breaks accuracy is the combination of accent with noise, reverb or compression, because the model loses the acoustic detail it needs to disambiguate. Fix the recording conditions and the accent stops being the limiting factor.

Should I clean up audio before transcribing or just upload the raw file?

Clean it up, but gently. Normalising levels and applying light noise reduction reliably helps, while aggressive processing hurts more than the original noise did. The safest approach is to transcribe both the raw and cleaned versions of a two minute sample, compare them, and only then process the full file with the settings that won.

What is the best way to handle two people talking at the same time?

Record them on separate tracks whenever you can, then transcribe each track independently and merge by timestamp. If you only have a mixed recording, accept that overlapping sections will need manual typing and focus your effort there. No current transcription tool reliably separates simultaneous speech from a single mixed channel.

How do I stop names and technical terms being transcribed incorrectly?

Supply them in advance through a custom vocabulary or keyword hints field before you transcribe. If your tool does not offer one, run the transcript once, note the consistent misspelling the model produced for each term, and use find and replace to fix every instance at once. Models make the same substitution repeatedly, so a handful of replacements usually clears the whole file.

Difficult audio rewards preparation far more than it rewards tool switching. Spend five minutes normalising levels, applying light noise reduction, choosing the right language variant and writing a list of names, and you will get a better transcript from almost any tool than you would from the best tool fed a raw noisy file. If you find yourself repeatedly fixing the same recording problem, change how you record rather than how you process; for background on capturing clean source audio, our tape script guide for audio producers is a useful next read.