Intron, a specialised technology company, has launched the newest version of its flagship AI voice model, “Sahara v2”, which can now understand over 500 distinct African accents, covering up to 57 languages.
- •New additions include: Hausa, Swahili, Zulu, Yoruba, Kinyarwanda, Igbo, Xhosa, African French.
- •The platform was trained on over 14 million audio clips, spanning 50,000+ hours, from 40,000+ speakers representing 64+ African and diaspora languages over 30 African Countries
- •The release also includes a bilingual Swahili-English ASR (Automatic Speech Recognition), the first of its kind, developed with Penda Health, Kenya.
“In our testing, accuracy was excellent on several Southern African accents and APIs were robust with 99%+ success rates.” said Sarah Morris, CPO at Audere.
Global speech AI models have shown that they were not designed for the nuances inherent to African languages: the tonal richness, code-switching patterns, and accent diversity of African speech.
According to Intron, Sahara v2 was evaluated on the basis of “word-error rate" (WER), calculating the percentage of errors relative to the total words in a reference transcription. Where the lower the output value, the more accurate the model, with results showing that Sahara v2 (yellow) consistently maintains the lowest bars across all languages.
