Code-Switching: When One Sentence Holds Three Languages

"I told them say the deadline don pass, but they still want am delivered by Friday."

That is one sentence, in three registers, and nobody who speaks it thinks of it as unusual. It is how a great many Nigerians talk at work, at home and on the phone.

It is also the single hardest thing you can hand a transcription system.

Why it is hard, mechanically

Most speech systems decide on a language, load the expectations that go with it, and interpret sound through them. Those expectations are doing a lot of work — they are how a system resolves an ambiguous sound into a plausible word.

When the speaker switches languages mid-sentence, those expectations become actively wrong. The system does not usually notice and stop. It carries on confidently, mapping unfamiliar sounds onto the nearest familiar words, and produces fluent output that means nothing.

That failure mode is worse than an obvious error, because it looks like a transcript.

Switching is rule-governed, not random

The common assumption is that code-switching is sloppiness. It is not — it follows patterns, and the patterns are informative.

People switch to mark emphasis, to signal intimacy or distance, to quote someone, to soften a refusal, to be funny, or because a particular concept simply lives in one language. A negotiation that moves into Pidgin at the difficult moment is telling you something about the negotiation.

For research and for quality work, that means the switch itself is data, not noise to be cleaned away.

What helps

Tell the system the truth about the recording. This is the biggest single factor. If a conversation is Yoruba and English, saying so gives it a chance; setting it to English alone does not. The languages page lists what is supported and where support is partial.

Expect to review the switch points. They are where errors cluster. Skimming a transcript with the audio open and paying particular attention to the transitions catches most of what went wrong.

Do not "correct" the speech. A transcript that silently renders Pidgin into standard English has thrown away the thing that was said. If you are quoting, quote what was spoken.

Where it shows up most

Voice notes, because they are recorded the way people speak rather than the way they write. Focus groups, once participants relax. Interviews, at exactly the moment somebody says something they mean. Sermons and negotiations, where the switch marks the point.

There is more on the underlying difficulty in why transcription tools mangle Nigerian English and on Pidgin specifically in Nigerian Pidgin transcription.

The honest position

Code-switched speech is harder than single-language speech, for everyone, and any tool that tells you otherwise is selling something. What matters is whether a system treats it as the normal case to be handled or an edge case to be apologised for.

A free account includes transcription minutes each month with no card. Sign up and test it on your most code-switched recording, because that is the one that will tell you something.