One numeral take · six functions

License one numeral recording —
it does the work of six datasets.

A native speaker says a number in Bantu, then crosses to the English value — in one take. Your model can't count in Bantu; this is the structured proof, and the fix.

benchmarked numeral takes

datasets per take

100%

consented · de-identified

Worked example · TSN · None

Hear one take do the work.

The highlight swaps at the stored boundary. Play the Bantu numeral and the English value separately — that's the segmentation a number-ASR model needs. { } machine record shows it's real structured data.

Numeral code-switch · boundary stored

bongwe, bobedi, boraro, bone, botlhano, borataro, bosupa, borobabobedi, borobabongwe, lesome, lesome le bongwe, lesome le bobedi, lesome le boraro, lesome le bone, lesome le botlhano, lesome le borataro, lesome le bosupa, lesome le borobabobedi, lesome le borobabongwe, masome a mabedi

means one to twenty

{
  "iso": "tsn",
  "n_value": null,
  "kind": "bantu_english",
  "switch_ms": 53556,
  "duration_ms": 68140,
  "segments": [
    {
      "label": "bantu",
      "start_ms": 0,
      "end_ms": 53556
    },
    {
      "label": "english",
      "start_ms": 53556,
      "end_ms": 68140
    }
  ],
  "speaker": "Speaker BEM-B",
  "consent": "granted",
  "deidentified": true
}

Six functions

What one numeral take is worth.

Number-word ground truth

The exact Bantu words for a value — recorded, native-verified. Your model stops hallucinating Bantu numbers.

Code-switch number ASR & language-ID

Bantu→English at the value, with the stored switch point — training + eval for how people actually say numbers.

Bantu-accented English numbers

The English half, native-accented — accent-robust number recognition for Bantu-accented speech.

FSI alignment target

Each Bantu numeral decomposes into its language's Full Syllable Inventory — the units an aligner needs.

Number pronunciation / TTS

Consented native pronunciations keyed to value — a clean basis for number TTS and voice.

Turnkey number ASR benchmark

The English value is isolated and machine-transcribable against a gold answer — a self-grading eval set.

Function 06, proven · Google Cloud Speech-to-Text

The benchmark grades itself.

We isolated the English value and gave it to a frontier ASR (Google Cloud Speech-to-Text). It drops the marker, mangles the value, or normalizes it wrong. The gap is the work — gold answer key built in.

value mangled

goldmeans four thousand nine hundred ninety nine in money

Cloud STTMeans 4,999 in money.

content recall 14% · model confidence 70%

value mangled

goldmeans nine hundred ninety nine point nine nine in money

Cloud STTMeans 999.99 in money.

content recall 12% · model confidence 66%

value mangled

goldmeans one to twenty

Cloud STTMeans 1 to 20.

content recall 0% · model confidence 90%

value mangled

goldmeans one hundred twenty three thousand four hundred fifty six

Cloud STTMeans 123,456.

content recall 0% · model confidence 65%

value mangled

goldmeans nine hundred ninety nine thousand nine hundred ninety nine

Cloud STTMeans 999,999.

content recall 0% · model confidence 45%

value mangled

goldmeans nine eight seven six five four three two one

Cloud STTMeans 9, 8, 7 6, 5, 4, 3 2 1.

content recall 0% · model confidence 93%

Part of the BantuNomics family.

One consented recording = six datasets — for numbers, tone, and clinical health alike. License the program, not a dataset.

Talk to us →

License one numeral recording —it does the work of six datasets.

Hear one take do the work.

What one numeral take is worth.

The benchmark grades itself.

Part of the BantuNomics family.

Start where you are.

License one numeral recording —
it does the work of six datasets.