Verbatimly

Preparing a recording for transcription

Updated 15 November 2026

Automatic speech recognition is good at clean speech and increasingly bad at everything else. When a transcript comes back wrong, the cause is usually one of three conditions in the recording — and none of them can be fixed after the fact.

One: overlapping speakers

When two people talk at once, the model receives a single mixed signal and has to produce a single stream of text. It will pick one voice, drop the other, or invent a blend of both that neither person said.

What helps, in order of how well it works:

Two: music and background noise

Background music is the second most common cause, and it is more damaging than most people expect. A steady, quiet bed of music is fine. Music with vocals in it is not: the model hears two voices and tries to transcribe both, producing text that appears to hallucinate.

Other background sounds follow the same logic. An air conditioner, traffic, or a fan is tolerable if it is quiet relative to the voice. A television in the next room is not, because it is speech.

If you cannot record without background, the practical options are to get the microphone closer to the speaker (which raises the voice relative to everything else) or to clean the audio in an editor before transcribing.

Three: distance from the microphone

A recording made from across a room is quieter and more reverberant, and both hurt. Reverberation smears the ends of words into each other; a distant microphone lets every other noise in the room compete with the voice.

The rule of thumb: the microphone should be within a forearm's length of the mouth. A laptop's built-in microphone at normal sitting distance is borderline — fine in a quiet room, not fine in a café.

One specific trap: a headset microphone placed too close and directly in front of the mouth produces plosives. P and b sounds become a burst of low-frequency noise that reads as an unintelligible syllable. Angle the microphone slightly to the side of the mouth rather than in front of it.

Things that matter less than people think

Accents. Modern models handle most accents well. An accent produces more errors than clean native speech, but far fewer than music does. If the transcript reads as a plausible rendition with a few wrong words, that is the accent and not much can be done about it. If it reads as gibberish, look at the audio conditions instead.

File format and bitrate. Audio that has already been compressed to MP3 at a low bitrate is slightly worse than WAV, but the difference is small compared to microphone placement. Converting a WAV to MP3 to save upload time is not the thing that will ruin your transcript.

Speaking speed. Fast speech crowds words together and causes errors at word boundaries. It is worth asking a fast speaker to slow down, but it is a smaller effect than the three conditions above.

If the recording already exists

You cannot undo a mixed room recording, but two things are worth trying before you accept a poor transcript:

  1. Trim the silent stretches. Long pauses sometimes cause the model to restart a segment badly.
  2. Retry the file. Transcription is not perfectly deterministic; a second run occasionally produces a materially different result on marginal audio.

If both fail, the transcript is as good as that recording allows. In that case, read along with the audio and correct the transcript by hand for the sections that matter, rather than re-running it repeatedly.

Recording-for-transcription checklist

Ten seconds of setup removes most of the work you would otherwise do correcting the transcript afterwards.

Transcribe a recording →