Preparing Interview Audio for Transcription: Format, Bitrate, Cleanup

Whether you transcribe with an AI service, a human transcriptionist, or your own two ears, one fact holds: transcription accuracy is decided before the transcript starts, by the audio you feed in. Services quote error rates against clean audio; hand them a quiet, boomy phone recording with two voices spread oddly across stereo channels and the error rate doubles or triples. The good news is that most of the gap is closed by a five-minute preparation pass — format conversion, mono downmix, and loudness normalization — that anyone can run in a browser.

This guide covers what actually affects speech-recognition and human-transcription accuracy, and the exact prep pipeline for interview audio.

What transcription engines are sensitive to

Modern speech-to-text models are remarkably robust to accents and vocabulary, but they remain sensitive to a handful of signal-level factors:

• Level. Audio that is too quiet loses consonants into the noise floor; audio that clips distorts them. Both produce misrecognitions.
• Consistency between speakers. If the interviewer is 12 dB louder than the subject — the classic phone-on-the-table scenario — the quieter voice transcribes far worse.
• Channel layout. Some engines process one channel or downmix unpredictably. A voice recorded mostly into the left channel can be degraded or half-lost.
• Codec damage. Very low bitrates smear the high-frequency detail that distinguishes "s" from "f" and "t" from "k."

Every one of these is fixable after the fact except severe clipping — which is why prep matters.

Step 1: Get everything into one clean format

Interviews arrive in whatever the recording device chose: M4A from iPhone Voice Memos, WAV from field recorders, sometimes odd container formats from call-recording apps. Standardize on MP3 — every transcription service accepts it, and it keeps your project folder sane. Convert Apple recordings with an M4A to MP3 converter and recorder files with a WAV to MP3 converter. If your recordings are Voice Memos specifically, the Voice Memos FAQ covers where iOS hides the files.

Bitrate: there is a floor

Transcription does not need audiophile bitrates, but it punishes starvation. Keep speech at 128 kbps or above when converting; if a source file is already at 64 kbps, transcribe it as-is rather than re-encoding, since re-encoding cannot restore lost detail. The bitrate guide explains why low bitrates hit consonant intelligibility first — the same frequencies transcription engines lean on.

Sample rate: leave it alone

One setting that does not need your attention is sample rate. Speech recognition engines internally resample everything to their model's native rate anyway, so whether your recording is 44.1 kHz or 48 kHz makes no measurable difference to accuracy — the sample rate FAQ explains why both comfortably cover the human voice. Keep whatever the recorder produced and spend your effort on the steps that follow, which actually move the error rate.

Step 2: Downmix to mono

Unless you deliberately recorded each speaker on a separate channel for split-channel transcription (some services support this and it genuinely helps speaker labeling), stereo is a liability for interview audio. A phone on a table records the room into both channels with random imbalance; a downmix to mono averages the channels and presents the engine with one coherent signal. Run files through a stereo to mono converter — it also halves the file size, which matters when you are uploading hours of tape. The mono vs stereo FAQ covers the rule in general.

The one exception: if you know your recorder placed the interviewer on the left channel and the subject on the right, check whether your transcription service offers per-channel processing before you downmix — that layout is worth preserving.

Step 3: Normalize the loudness

This is the highest-value step and the most skipped. A loudness normalization pass raises quiet recordings to a standard level and evens out the gap between a loud interviewer and a soft-spoken subject. For transcription prep you are not chasing broadcast polish — you are making sure every word sits well above the noise floor and well below clipping. The normalization FAQ explains the difference between peak and loudness normalization; loudness normalization is the one you want for speech.

What prep cannot fix

Honesty about the limits saves wasted effort:

• Heavy background noise. Normalization raises the cafe noise along with the voices. Dedicated noise reduction can help, but a noisy recording will always transcribe worse than a quiet one.
• Crosstalk. Two people speaking simultaneously defeats humans and machines alike. No file-level fix exists.
• Clipping. Distortion from recording too hot is baked in. If a recording clips throughout, budget for extra correction time.

These are recording-stage problems — the lesson for next time is a quiet room, the device closer to the quieter speaker, and levels checked in the first minute.

The complete pre-flight checklist

• Convert every recording to MP3, 128 kbps or higher.
• Downmix to mono (unless using deliberate split-channel recording).
• Normalize loudness across all files.
• Trim pre-interview chatter with a trim tool — many services bill by the audio minute, and five minutes of setup talk per interview adds up across a project.
• Name files consistently: date, participant, session number.

The payoff

On clean, normalized, mono speech, good AI transcription now lands in the 92 to 98 percent accuracy range; on unprepped phone audio it can fall to 80 percent or worse. That difference sounds abstract until you are the person fixing it: at roughly 150 words per minute of speech, each percentage point of error is about 90 wrong words per hour of tape to find and correct by hand. Five minutes of browser-based prep per interview routinely saves an hour of transcript cleanup — the best exchange rate in the entire research workflow.