Meeting transcription that understands Nigerian English and Pidgin

If you have transcribed a Nigerian meeting with a global AI tool, you have seen the output: directors' names mangled, Pidgin rendered as phonetic guesses, and formal terms like "ordinary resolution" turned into something unrecognisable. The problem is not your microphone. It is who the models were trained on.

The accuracy gap nobody talks about honestly

The big transcription tools quote 90–95% accuracy — measured on clean American speech. Independent testing tells a different story for everyone else: accuracy drops to roughly60–70% once accents, code-switching, and crosstalk enter the room. For Nigerian meetings — where speakers move between formal English, Nigerian English, Pidgin, and vernacular in a single sentence — that gap is the product.

Why the models fail on Nigerian speech

Three reasons compound:

• Training data. Speech models are trained predominantly on American and British audio. Nigerian English has its own vowel patterns, intonation, and vocabulary that the models have barely heard.
• Code-switching. A single Nigerian sentence can move between English, Pidgin, and vernacular. Models tuned to one variety break at every switch.
• Names and places. Adeyemi, Ogunlesi, Nwosu, Ikeja, Onitsha — phonetic guesses produce wrong names in the one place accuracy matters most: who said what.

How CosecScribe handles it

CosecScribe was built in Lagos, for Nigerian meetings. Two things make the difference:

Transcription tuned for Nigerian speech. Our pipeline uses AssemblyAI's current Universal streaming models — the strongest available for accented speech — combined with AI post-processing that explicitly understands Nigerian English expressions and Pidgin. When someone says "abeg make we hold am there", the transcript and the summary reflect what was actually meant.

Domain context. The AI is instructed that it is analysing a Nigerian business meeting — corporate governance, client matters, formal registers — so "quorum", "proxies", and "ordinary resolution" come out as the legal terms they are, not phonetic approximations.

Practical tips for better transcripts in any tool

• Place the microphone centrally — one device per eight people, roughly.
• Ask side conversations to pause; crosstalk is the hardest case for every model.
• For virtual meetings, the notetaker captures the platform audio directly — quality is near-perfect regardless of accents.
• For in-person meetings, a phone on the table outperforms a laptop across the room.
• Speak in full sentences where possible; fragments are harder for every model.

Test it on a real Nigerian meeting

Meeting capture is free for every CosecScribe user — online with the notetaker, or in person from your device. Run it on a meeting with real Nigerian speech and compare the output to whatever you are using today.