BantuNomics My BantuNomics

The Flat Text Problem

Flat text is writing without the sound layer.

The Flat Text Problem is the failure of ordinary written Bantu text to carry all the sound information needed to determine meaning.

A simple definition.

Flat text is the written form after tone, vowel length, downstep, phrase boundary, rhythm, and voice quality have been stripped away or left unmarked.

In many Bantu languages, standard spelling is useful and efficient for native readers. But it is efficient because it assumes the reader already knows the language and can hear what the page does not show. The reader supplies the missing sound layer from memory, grammar, context, and native intuition.

AI systems trained mostly on text do not have that native sound layer. They receive the spelling, but not the tonal and phonetic evidence that can separate one meaning from another. The model therefore treats distinct spoken forms as if they were the same flat string.

Why this is a challenge for AI.

Bantu languages have been spoken for millennia. Their modern writing systems are comparatively recent and were not designed to encode the full sound structure of the language. Standard orthography often leaves tone unmarked and does not reliably expose vowel length, downstep, phrase edge, or other acoustic cues.

For a human speaker, the page can still work. For ASR, TTS, translation, search, health language, education, and model evaluation, the omission becomes structural. The system is asked to understand the language while being denied part of the signal that decides meaning.

Why labs miss it

Four facts hiding in plain sight.

Tone is unwritten

Standard Bantu orthography normally does not mark the pitch pattern that distinguishes meaning.

Tokenizers see letters

BPE optimizes character frequency. It does not discover the syllable as the meaning-bearing sound unit.

The phonology lives off-web

A century of scholarship and native knowledge is not sitting cleanly in CommonCrawl.

Audio is canon

For tone, vowel length, phrase edge, downstep, and voice quality, the native audio is the evidence.

Ulebomba is the evidence.

In standard Bemba writing, ulebomba appears as one written form. In speech, the same spelling can resolve into ten distinct readings. The difference is not in the letters. It is in the sound layer: tone, vowel length, phrase edge, and related cues.

This is why the example matters. It does not merely illustrate ambiguity. It shows the central failure mode of text-only Bantu AI: the written token is too flat to carry the meaning space the spoken language actually uses.

Audible diagnostic

One spelling. Ten meanings. Hear the missing layer.

In standard Bemba writing, ulebomba is one eight-letter form. In speech, it can resolve into ten distinct readings. The differences live in tone, vowel length, downstep, phrase boundary, and voice quality โ€” play them below.

01

You are working

present statement

02

You are getting wet

present statement

03

You should work

modal statement

04

Are you working?

yes/no question

05

Who is working

relative reading

06

You are really working

emphatic reading

07

Is the one working?

question with focus

08

Is the one getting wet?

question with lexical shift

09

Are you getting wet?

direct question

10

You should be getting wet

modal progressive

What follows from FTP

More scraped text will not restore a layer the writing system never encoded.

A tone-aware Bantu system needs Full Syllable Inventories, syllable-level alignment, consented native audio, and a way to represent the missing sound evidence. That is where A12 enters the product line: it turns audio plus spelling into layered, auditable structure.