Audio to text

Audio is where accuracy is decided

Transcription quality tracks recording quality almost exactly, and that is worth knowing before you record rather than after. The things that cost you accuracy, roughly in order of how much damage they do:

• ·Two people talking at once. Overlapping speech is the single hardest case, and no tool solves it well.
• ·Distance from the microphone. A phone on a table across the room picks up more room than voice.
• ·Echo. Hard walls and bare floors smear every word into the next one.
• ·Background noise. Cafés, traffic and air conditioning all eat consonants.
• ·Specialist vocabulary. Names, drugs, case numbers and jargon are guessed from context, and the context is often not enough.

None of that requires equipment to fix. Moving the phone closer, closing a window and asking people not to talk over each other will do more for your transcript than any setting in any product.

Audio formats it accepts

MP3, WAV, M4A, AAC, FLAC, OGG, Opus, WMA, AIFF, AMR and CAF — which covers what phones, voice recorders, meeting software and podcast rigs produce. Compressed or uncompressed makes no practical difference to the result: a 128 kbps MP3 of a clear conversation transcribes about as well as the WAV it came from.

Who is speaking

When a recording has more than one voice in it, speech is grouped by speaker and each turn is labelled, so an interview comes back as an exchange rather than one undifferentiated block. The labels start out generic and you rename them to the real names in one place each.

This is worked out from the sound of each voice, so it is very good at hearing that the speaker changed and imperfect at deciding how many distinct people there are — especially with similar voices or heavy crosstalk. It handles up to eight speakers.