Why transcripts get words wrong

A transcript error is a word in the text that doesn't match what the person said. Tools make them for several reasons, and most trace back to the sound, the speaker or the vocabulary.

The voice itself

Speech varies a lot. Wikipedia lists the ways: "Vocalizations vary in terms of accent, pronunciation, articulation, roughness, dialect, nasality, pitch, volume, and speed"¹.

A tool learns from examples. Voices far from those examples are harder for it. The same page notes that a system meant for any speaker is "more difficult" to build than one meant for a single speaker¹.

Natural talk is messier than reading aloud. With spontaneous speech, "recognition must deal with disfluencies such as" um and uh, false starts, coughing and laughter¹.

The recording

Sound quality matters before the software even starts. Wikipedia says that "Speech is distorted by background noise, echoes, and recording characteristics"¹. Google's guidance says that "Excessive background noise and echoes may reduce accuracy, especially if a lossy codec is also used"².

A lossy codec is a file format that throws away some sound detail to make the file smaller. The page on recording tips covers what you can change.

Names, jargon and acronyms

Every tool has a vocabulary. Google's documentation says that "terms and proper names that are out-of-vocabulary will not be recognized"². A small supplier's name or a product code is a typical example.

Amazon makes the same point. If media contains "domain-specific or non-standard terms, such as brand names, acronyms," its tool may not capture them correctly³.

Some makers offer a fix. Amazon describes custom vocabularies as "designed to tune" recognition for specific words³. Other tools call this a word list, glossary or hint. Look for it in the maker's documentation.

People talking at once

Crosstalk is two or more people speaking at the same time. Google's guidance says that "Multiple people talking at the same time, or at different volumes may be interpreted as background noise and ignored"².

That means words can go missing, not only come out wrong. A quiet speaker at the far end of a table is at risk in the same way.

What this means for checking

Errors don't always look like errors. A wrong word that's a real word reads smoothly, so a quick skim won't catch it. Names, numbers, dates and quoted terms deserve a second look against the audio before anyone relies on the text.

Start with the parts that would cause trouble if wrong: an amount, a deadline, a customer's name, a decision. Play those few seconds of audio and compare. It takes far less time than reading the whole transcript against the recording, and it targets the spots where the tool is most likely to have guessed.

Also watch for gaps. As noted above, speech that overlaps can be dropped, so a transcript that reads well may still be missing a sentence. If a passage feels thin, listen to that stretch of audio again.