Not all noise is equally difficult
Constant noise is the easy case. Air conditioning, traffic hum, a fan, road noise from a car, a steady hiss from old equipment: these sit at a predictable level behind the speech, and modern speech recognition copes with them far better than most people expect.
Other voices are the hard case. A café, an open-plan office or a crowded room puts speech behind speech, and nothing distinguishes the conversation you want from the one at the next table when both are just people talking. Heavy echo is similarly difficult, because a reflective room smears each word into the next.
Why cleaning the file up first usually backfires
The instinct is to run noise reduction before transcribing, and it is usually a mistake. Aggressive filtering removes the quiet high-frequency detail that distinguishes similar consonants, so the recording sounds cleaner to you while giving the transcription engine less to work with than the noisy original did.
If you want to try it, test on a short clip and compare the transcripts rather than trusting your ears. Light, conservative removal of a constant hum is reasonable. Anything heavier, and you are trading real information for the impression of quality.
Frequently asked questions
Will noise reduction improve my transcript?
Usually not, and often the opposite. Heavy filtering strips detail the engine relies on. Test any cleanup on a short section and compare transcripts before applying it to the whole recording.
What can I do about people talking in the background?
Very little after the fact, because background speech is the same kind of sound as the speech you want. Expect gaps in those passages and mark them honestly rather than guessing what was said.
Is it worth transcribing a genuinely bad recording?
Often yes, as a finding aid. Even a rough transcript makes hours of audio searchable, so you can locate the moment you need and listen to that part properly rather than scrubbing through everything.