Why Most AI Tools Can't Transcribe Kikuyu (and What We Do Instead)

This comes up often enough that it's worth explaining plainly, without the usual vague "AI doesn't support every language yet" hand-wave: Kikuyu breaks automated transcription for a specific, real reason, not a vague one.

The actual reason: no ASR model exists

Automatic speech recognition (ASR) - the step that turns spoken audio into text - depends on a model trained on a large amount of transcribed audio in that specific language.

For Kikuyu, that training data largely doesn't exist in a form anyone has built a production-grade ASR model from yet.

This isn't unique to Kikuyu; it's the same reason thousands of languages worldwide have no usable AI transcription today, even ones with millions of speakers - model training needs data at a scale that simply hasn't been assembled for most of the world's languages.

Machine translation and voice synthesis, further down the pipeline, generally depend on that first transcription step being accurate.

If it's missing entirely, the rest of an "AI-only" pipeline has nothing real to work from.

What actually works: a person, not a workaround model

The practical fix isn't a clever technical trick - it's a real person doing the transcription step by ear, the same way professional transcription worked before ASR existed at all for any language.

Once that transcript exists as real text, translation and voice synthesis can run the same AI-assisted pipeline used for every other language Kauli supports - the manual step is specifically the transcription stage, not the whole process.

What this means for cost and turnaround

Manual transcription takes real human time in a way an automated first pass doesn't, which is the actual, disclosed reason Kikuyu orders cost somewhat more per audio-minute than Swahili or English orders - it's covering real paid labor, not an arbitrary "less common language" surcharge.

Turnaround accounts for the same real difference - an order's delivery estimate factors in the actual service level and language involved, not a flat number that pretends every language costs the same to process.

The upside worth remembering: this doesn't make Kikuyu content lower-quality once it's done - it means the first step took a person instead of a model, and the same human-reviewed standard applies to everything after that, same as every other order.