Position · why tone
The spoken word is the ground truth.
Bantu has been spoken for roughly three thousand years and written for about one hundred. The writing is the recent, lossy part. For an AI system, that single fact changes what counts as truth — and what you can and cannot learn from text.
The asymmetry no one prices in
Put the two timelines side by side. A Bantu language has carried meaning in speech for on the order of 3,000 years. It has been written down for roughly 80–100. Writing covers something like the last 3% of the language's life. And those orthographies were devised recently — largely from the outside, for practical literacy — and optimized for a reader who already knows the language and can fill in the blanks. The page was never meant to carry the whole signal. It was meant to remind a native of it.
What the page leaves out, on purpose
Standard Bantu orthography does not mark tone, and usually does not mark vowel length.
That is not an oversight; it is efficient compression for an in-group reader who restores
those layers from context. The Bemba word ulebomba is one written form that
resolves, in speech, into ten distinct meanings. The distinctions are real, grammatical,
and entirely unwritten — they live in pitch, length, downstep, and phrase edge. A reader
who speaks the language never notices the gap. Everyone else — and every model — gets the
compressed string with the decoder missing.
The inversion
So the natural hierarchy is the reverse of what the text-first world assumes. Speech is the canonical artifact; the writing is a recent, lossy index into it. A model trained only on the text is trained on the century-old shadow and mistakes it for the substance. The tone and length it needs were never in its training data — not because the data was thin, but because the medium drops them by design.
Grounded by meaning
Across those three thousand years, the spelling was never the invariant. Orthographies are young, regional, and reformed. The stable, load-bearing thing is the meaning ↔ sound pairing: the meaning selects which realization a speaker produces, and the sound carries it. The written form is an optional, late-arriving, lossy pointer to that pair. So the irreducible unit of truth is not the word on the page — it is (meaning, audio). That is why everything we capture is keyed to a meaning, not merely to a spelling: fix the sense, let a native speaker realize it, and record the truth.
Why scale cannot close it
This is a modality gap, not a volume gap — the most expensive thing to get wrong. You cannot learn from text what text systematically leaves out. Scrape the entire written web of a Bantu language and the tone is still absent — not sparse, absent. More text multiplies the same compressed signal; it never recovers the layer the orthography never wrote down. No amount of scale supplies a dimension that was never encoded.
Why the frontier does not see it
The blindness is a frame error, not a knowledge gap. Western languages have expansive, open syllable spaces, so the syllable was never a clean operating unit there; NLP moved to letters and sub-word tokens. That intuition gets universalized — and it is silently wrong for Bantu, whose syllable inventory is finite and closed and brackets the whole language. A sub-word tokenizer is built, by construction, to ignore exactly the structure that — for Bantu — is the alphabet. So the gap is unseen, not refuted.
The remedy
The work, then, is to capture the spoken word grounded by meaning, and make the implicit explicit and machine-consumable: meaning-disambiguated native audio for homographs (tonal ground truth); the finite Full Syllable Inventory to decompose any word into its tone-bearing syllables (459 assembled — the first complete collection in history); the time-segmented Bantu→English code-switch corpus; and a measured benchmark of how a frontier ASR handles the accent. It is the three-thousand-year-old reality, captured and externalized — not the hundred-year-old approximation of it.
Hear the argument
Don't take the claim — play it.
One spelling, many meanings, in real consented native audio; the syllable as the tone-bearing unit; and a frontier-ASR accent baseline. The spoken word, grounded by meaning, made testable.