Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-07-26 · source-backed
Paul Azunre released twenty-one monolingual and five multilingual w2v-BERT 2.0 ASR base models spanning varieties across Ghana, Sierra Leone, Nigeria, Senegal, Kenya, and Zimbabwe, languages with roughly a hundred million first-language speakers (arXiv 2607.21540). An annealed multi-step learning-rate schedule plus language conditioning via one-hot language identity prefixed to acoustic features brings the five multilingual families to 10-13% average WER, closing most of the gap to monolingual. Apache-2.0 attribution-only on the Hugging Face KhayaAI org, so commercial fine-tuning is permitted. Caveat: training data is primarily religious text, which constrains domain coverage hard.
Each link below shares sources, entities, or timing with this story.
Shared entities / Shared topic / Earlier coverage
Both cover Apache, WER; overlapping topics (apache-2, average, model); earlier Apache coverage from 2026-06-08.
Shared entity: Apache / Shared topic / Earlier coverage / Tension
Both cover Apache; overlapping topics (apache-2, model); earlier Apache coverage from 2026-06-12.
Shared entities / Earlier coverage
Both cover ASR, WER; earlier ASR coverage from 2026-07-20.
Both cover Apache, ASR; earlier Apache coverage from 2026-04-30.
Both cover Apache, WER; earlier Apache coverage from 2026-03-17.
Shared entity: Apache / Shared topic / Earlier coverage
Both cover Apache; overlapping topics (apache-2, model); earlier Apache coverage from 2026-07-25.
Shared entity: ASR / Same source domain / Earlier coverage
Both cover ASR; reported by the same outlet (arxiv.org); earlier ASR coverage from 2026-07-21.
Shared entity: Apache / Same source domain / Earlier coverage
Both cover Apache; reported by the same outlet (arxiv.org); earlier Apache coverage from 2026-07-16.