BantuNomics My BantuNomics

Benchmark · frontier-ASR baseline

How a frontier ASR hears Bantu-accented English.

Every code-switch take is time-segmented at the Bantu→English boundary, so the English half is isolated, transcribed by a frontier ASR (Google Cloud Speech-to-Text), and scored against the gold reference. The result is a measured, reference-backed accent-robustness benchmark — run your own model on the same audio and compare to this baseline.

227Takes scored
2Languages
81%Content recall
69%“means” misheard

The systematic failure mode

The accent breaks function words before it breaks meaning.

The substantive English is mostly recovered (content recall 81%), but the discourse marker means is mis-transcribed in 69% of takes — heard as “men's”, “main”, even “Main Street”. A high-frequency word a frontier ASR gets wrong almost every time on this accent. That is exactly the gap an accent benchmark is built to surface.

Listen + compare

Gold reference vs the ASR hypothesis.

Play the English segment (what the STT was given), then read the gold reference against Cloud STT's transcript.

akanya nga fwaya ukuya kucimbushu “means” misheard
akanya nga fwaya ukuya kucimbushu
means she will defecate when she feels like going to the toilet

Audio pending.

Gold refmeans she will defecate when she feels like going to the toilet
Cloud STTmen's shoes defecate when she feels like going to the toilet
confidence 0.84 content recall 89% WER 0.30 Speaker BEM-A
akanya nga fwaya ukuya kucimbushu not detected as English
akanya nga fwaya ukuya kucimbushu
means he will defecate when he feels like going to the toilet

Audio pending.

Gold refmeans he will defecate when he feels like going to the toilet
Cloud STTyou defecate
confidence 0.48 content recall 11% WER 0.90 Speaker BEM-A
Akabomba ku ofesi namumushi nga afika not detected as English
Akabomba ku ofesi namumushi nga afika
means he will work at the office tomorrow when he arrives

Audio pending.

Gold refmeans he will work at the office tomorrow when he arrives
Cloud STTmemes
confidence 0.45 content recall 0% WER 1.00 Speaker BEM-A
Alelanda ne mfumu nga bafika ku ng'anda not detected as English
Alelanda ne mfumu nga bafika ku ng'anda
means he was speaking with the chief when they arrived at the house

Audio pending.

Gold refmeans he was speaking with the chief when they arrived at the house
Cloud STTmeans
confidence 0.59 content recall 0% WER 1.00 Speaker BEM-A
Abantu babomba ku ofesi lyonse not detected as English
Abantu babomba ku ofesi lyonse
means the people work at the office every day

Audio pending.

Gold refmeans the people work at the office every day
Cloud STTmeans
confidence 0.55 content recall 0% WER 1.00 Speaker BEM-A
Bafwile mu nshita ya nsala ya myaka iyibili content review
Bafwile mu nshita ya nsala ya myaka iyibili
means they died in the famine of two years ago

Audio pending.

Gold refmeans they died in the famine of two years ago
Cloud STTmeans
confidence 0.64 content recall 0% WER 1.00 Speaker BEM-A
akanya umunani content review
akanya umunani
means share the relish

Audio pending.

Gold refmeans share the relish
Cloud STTmeans share the radish
confidence 0.66 content recall 50% WER 0.50 Speaker BEM-A
ulya akanya! content review
ulya akanya!
means that one will face severe consequences

Audio pending.

Gold refmeans that one will face severe consequences
Cloud STTmeans that one who said severe consequences
confidence 0.60 content recall 67% WER 0.33 Speaker BEM-A
akanya kalelila content review
akanya kalelila
means the baby is crying

Audio pending.

Gold refmeans the baby is crying
Cloud STTmeans the babies crying
confidence 0.74 content recall 33% WER 0.67 Speaker BEM-A

Per-language intelligibility

Measured accent difficulty.

LanguageScoredContent recallDeviation WER“means” errNot-EnglishMean conf.
🇸🇿 Swati215 82% 1.14 72%7% 0.73
🇿🇲 Bemba12 46% 0.56 25%33% 0.66

The offer

A turnkey accent-robustness eval, with a frontier baseline.

Run your ASR on the same time-segmented audio, score against the gold references, and compare to this Cloud-STT baseline. It is verification/validation test data for Bantu-English speech — native, consented, reference-backed, and growing. Pull it from /api/v1/tone/asr-benchmark.